AIPUB归智期刊联盟
Mobile website
- Latest News
- Recommended
- Experimental Evaluation and Improvement of Bitcoin Network Node Detection Method
- Unsupervised Remote Sensing Cloud Occlusion Image Restoration Based on Temporal-spectral Domain Fusion and Temporal Self-attention Enhancement
- Adaptive Dual-branch Dense Pedestrian Detection with Improved Attention
- Monocular-camera-based Visual Localization for Photovoltaic Panels
- Survey on Causally Disentangled Representation Learning
- ResMobileNet: Primary-secondary Residual Dual-branch Network for Cloud and Cloud Shadow Segmentation More+
- Current Issue
- Online First
- Archive
- Click Rank
- Most Downloaded
- 综述文章
- Monographic Series
Article Search
Search by issue
Select AllDeselectExport
Display Method:
2025,34(7):1-13, DOI: 10.15888/j.cnki.csa.009942, CSTR: 32024.14.csa.009942
Abstract:
The key for artificial intelligence to fundamentally comprehend the world around us is to identify and disentangle hidden, potentially interpretable factors from observed low-level sensory data. Disentangled representation learning aims to extract these independent and interpretable latent variables from data, while causally disentangled representation learning further emphasizes the causal relationships among these latent variables, thereby more truly simulating the complexity of the real world. In light of the increasing importance of causal learning, this study provides a detailed and comprehensive introduction to relevant methods combining causal learning with disentangled representation learning, intending to support future development in disentangled representation learning. The study classifies causally disentangled representation learning based on commonly used causal learning methods, mainly discussing methods that integrate structural causal models with flow-based disentangled representation learning, as well as commonly used datasets and evaluation metrics. Furthermore, it analyzes practical applications of causally disentangled representation learning in image generation, 3D pose estimation, and unsupervised domain adaptation, and provides a forward-looking perspective on future research directions. This study reveals potential exploration paths for researchers and practitioners, promoting continuous development and innovation in this field.
The key for artificial intelligence to fundamentally comprehend the world around us is to identify and disentangle hidden, potentially interpretable factors from observed low-level sensory data. Disentangled representation learning aims to extract these independent and interpretable latent variables from data, while causally disentangled representation learning further emphasizes the causal relationships among these latent variables, thereby more truly simulating the complexity of the real world. In light of the increasing importance of causal learning, this study provides a detailed and comprehensive introduction to relevant methods combining causal learning with disentangled representation learning, intending to support future development in disentangled representation learning. The study classifies causally disentangled representation learning based on commonly used causal learning methods, mainly discussing methods that integrate structural causal models with flow-based disentangled representation learning, as well as commonly used datasets and evaluation metrics. Furthermore, it analyzes practical applications of causally disentangled representation learning in image generation, 3D pose estimation, and unsupervised domain adaptation, and provides a forward-looking perspective on future research directions. This study reveals potential exploration paths for researchers and practitioners, promoting continuous development and innovation in this field.
2025,34(7):14-22, DOI: 10.15888/j.cnki.csa.009940, CSTR: 32024.14.csa.009940
Abstract:
The cleaning and maintenance of photovoltaic (PV) panels are critical tasks in the operation of PV power stations. In this context, the PV cleaning shuttle systems equipped with robotic arms and PV cleaning terminals have emerged as an innovative solution. These systems require precise acquisition of PV panel poses, including the tilt angle and distance relative to the vehicle body. To address this issue, this study proposes a monocular-camera-based visual positioning method for PV panels. First, the YOLOv8-pose keypoint detection model is improved to enhance detection accuracy. The PSA mechanism is introduced to optimize the backbone network, while the DySample dynamic upsampling module and proposed ADown* downsampling module are used to strengthen the neck network. Next, by combining the improved YOLOv8-pose with the geometric features of the PV panels, a method for calculating the tilt angle and distance is proposed, thereby achieving the pose positioning of PV panels. Experimental results show that the proposed improved algorithm achieves a 26.2% and 20.1% increase in the accuracy of calculating tilt angles and distances compared to the original YOLOv8-pose, enabling more precise positioning of PV panels.
The cleaning and maintenance of photovoltaic (PV) panels are critical tasks in the operation of PV power stations. In this context, the PV cleaning shuttle systems equipped with robotic arms and PV cleaning terminals have emerged as an innovative solution. These systems require precise acquisition of PV panel poses, including the tilt angle and distance relative to the vehicle body. To address this issue, this study proposes a monocular-camera-based visual positioning method for PV panels. First, the YOLOv8-pose keypoint detection model is improved to enhance detection accuracy. The PSA mechanism is introduced to optimize the backbone network, while the DySample dynamic upsampling module and proposed ADown* downsampling module are used to strengthen the neck network. Next, by combining the improved YOLOv8-pose with the geometric features of the PV panels, a method for calculating the tilt angle and distance is proposed, thereby achieving the pose positioning of PV panels. Experimental results show that the proposed improved algorithm achieves a 26.2% and 20.1% increase in the accuracy of calculating tilt angles and distances compared to the original YOLOv8-pose, enabling more precise positioning of PV panels.
2025,34(7):23-36, DOI: 10.15888/j.cnki.csa.009878, CSTR: 32024.14.csa.009878
Abstract:
To address the low accuracy and high miss detection rates in pedestrian detection caused by complex background interference, this study proposes an adaptive dual-branch dense pedestrian detection algorithm, DACD-YOLO, incorporating improved attention mechanisms. First, the backbone network employs an adaptive dual-branch structure, which fuses different features through dynamic weighting while introducing depthwise separable convolution to reduce the computational cost, effectively mitigating the information loss present in traditional single-branch networks. Second, an adaptive vision center is proposed to enhance intra-layer feature extraction through dynamic optimization, with channel numbers reconfigured to balance accuracy and computational load. A coordinate dual-channel attention mechanism is then introduced, combining a heterogeneous convolution kernel design within a lightweight fusion module to reduce computational complexity and improve the capture of key features. Lastly, a dilation convolution detection head is utilized, fusing multi-scale features through convolutions with varying dilation rates, effectively enhancing feature extraction for small and occluded objects. Experimental results show that, compared to the original YOLOv8n, the proposed algorithm improves mAP@0.5 and mAP@0.5:0.95 by 2.3% and 2.2%, respectively, on the WiderPerson dataset, and by 3.5% and 4.6%, respectively, on the CrowdHuman dataset. The experiments demonstrate that the proposed algorithm significantly enhances accuracy in dense pedestrian detection compared to the original method.
To address the low accuracy and high miss detection rates in pedestrian detection caused by complex background interference, this study proposes an adaptive dual-branch dense pedestrian detection algorithm, DACD-YOLO, incorporating improved attention mechanisms. First, the backbone network employs an adaptive dual-branch structure, which fuses different features through dynamic weighting while introducing depthwise separable convolution to reduce the computational cost, effectively mitigating the information loss present in traditional single-branch networks. Second, an adaptive vision center is proposed to enhance intra-layer feature extraction through dynamic optimization, with channel numbers reconfigured to balance accuracy and computational load. A coordinate dual-channel attention mechanism is then introduced, combining a heterogeneous convolution kernel design within a lightweight fusion module to reduce computational complexity and improve the capture of key features. Lastly, a dilation convolution detection head is utilized, fusing multi-scale features through convolutions with varying dilation rates, effectively enhancing feature extraction for small and occluded objects. Experimental results show that, compared to the original YOLOv8n, the proposed algorithm improves mAP@0.5 and mAP@0.5:0.95 by 2.3% and 2.2%, respectively, on the WiderPerson dataset, and by 3.5% and 4.6%, respectively, on the CrowdHuman dataset. The experiments demonstrate that the proposed algorithm significantly enhances accuracy in dense pedestrian detection compared to the original method.
2025,34(7):37-47, DOI: 10.15888/j.cnki.csa.009939, CSTR: 32024.14.csa.009939
Abstract:
The rapid advancement of automation technology and robotics requires more precision in mobile robot path planning. To address the problems of poor convergence stability, low sample efficiency, and insufficient environmental adaptability in deep reinforcement learning for path planning in complex environments, this study proposes an enhanced path planning algorithm based on dueling double deep Q-network (R-D3QN). By constructing a dual-network architecture to decouple the action selection and value estimation processes, this method effectively alleviates the Q-value overestimation problem, thereby improving convergence stability. In addition, this method designs a temporal-prioritized experience replay mechanism combined with the spatiotemporal feature extraction capabilities of long short-term memory (LSTM) networks to improve sample utilization efficiency. Finally, this method proposes a multi-stage exploration strategy based on simulated annealing to balance exploration and exploitation, thereby enhancing environmental adaptability. Experimental results demonstrate that, compared to the traditional DQN algorithm, the R-D3QN algorithm achieves a 9.25% increase in average reward value, a 24.39% reduction in convergence iterations, and a 41.20% decrease in collision frequency in simple environments. In complex environments, it shows a 12.98% increase in average reward value, an 11.86% reduction in convergence iterations, and a 42.14% decrease in collision frequency. Furthermore, the effectiveness of the proposed algorithm is validated when compared with other enhanced DQN algorithms.
The rapid advancement of automation technology and robotics requires more precision in mobile robot path planning. To address the problems of poor convergence stability, low sample efficiency, and insufficient environmental adaptability in deep reinforcement learning for path planning in complex environments, this study proposes an enhanced path planning algorithm based on dueling double deep Q-network (R-D3QN). By constructing a dual-network architecture to decouple the action selection and value estimation processes, this method effectively alleviates the Q-value overestimation problem, thereby improving convergence stability. In addition, this method designs a temporal-prioritized experience replay mechanism combined with the spatiotemporal feature extraction capabilities of long short-term memory (LSTM) networks to improve sample utilization efficiency. Finally, this method proposes a multi-stage exploration strategy based on simulated annealing to balance exploration and exploitation, thereby enhancing environmental adaptability. Experimental results demonstrate that, compared to the traditional DQN algorithm, the R-D3QN algorithm achieves a 9.25% increase in average reward value, a 24.39% reduction in convergence iterations, and a 41.20% decrease in collision frequency in simple environments. In complex environments, it shows a 12.98% increase in average reward value, an 11.86% reduction in convergence iterations, and a 42.14% decrease in collision frequency. Furthermore, the effectiveness of the proposed algorithm is validated when compared with other enhanced DQN algorithms.
2025,34(7):48-58, DOI: 10.15888/j.cnki.csa.009888, CSTR: 32024.14.csa.009888
Abstract:
With the widespread use of Internet, an increasing number of users are inclined to share personal details and emotion on social platforms. These online text data often capture genuine expressions in various contexts, reflecting the users’ internal psychological traits and personality tendencies. In recent years, research on personality detection based on social media text has made significant progress. However, most researchers rely on unprocessed public datasets, which inevitably contain noise due to their collection process. In addition, there is an over-reliance on semantic features extracted by pre-trained models, with insufficient attention to psycholinguistic features. To address these issues, this study proposes a novel method for personality detection. First, a plug-and-play data cleaning module based on confident learning is used to remove noisy data and improve dataset quality. Second, multi-level psycholinguistic features are extracted to complement the semantic features of the text. The proposed method is evaluated on the public Kaggle MBTI dataset, with results showing that, compared to existing advanced methods, it achieves improvements of 5.48% in accuracy and 4.22% in F1-score.
With the widespread use of Internet, an increasing number of users are inclined to share personal details and emotion on social platforms. These online text data often capture genuine expressions in various contexts, reflecting the users’ internal psychological traits and personality tendencies. In recent years, research on personality detection based on social media text has made significant progress. However, most researchers rely on unprocessed public datasets, which inevitably contain noise due to their collection process. In addition, there is an over-reliance on semantic features extracted by pre-trained models, with insufficient attention to psycholinguistic features. To address these issues, this study proposes a novel method for personality detection. First, a plug-and-play data cleaning module based on confident learning is used to remove noisy data and improve dataset quality. Second, multi-level psycholinguistic features are extracted to complement the semantic features of the text. The proposed method is evaluated on the public Kaggle MBTI dataset, with results showing that, compared to existing advanced methods, it achieves improvements of 5.48% in accuracy and 4.22% in F1-score.
2025,34(7):59-71, DOI: 10.15888/j.cnki.csa.009908, CSTR: 32024.14.csa.009908
Abstract:
Cloud occlusion in optical remote sensing images is one of the core challenges in remote sensing data processing. To address the limitations of current cloud removal technologies in handling cloud edge information and preserving image details, a generative adversarial network (TGAN) based on temporal-spectral domain fusion and temporal self-attention enhancement is proposed. Through its two-stage modular design, TGAN simultaneously improves the quality of remote sensing image restoration and processing efficiency. In the first stage, the feature extraction module, based on a temporal self-attention mechanism, uses a linear expansion layer to capture temporal and spectral domain features, compensating for the limitations of traditional maximum pooling with a one-dimensional linear dimensionality reduction layer, thus enhancing the modeling capability of time-series positional features. A multi-head self-attention mechanism with a weight allocation strategy is designed to accurately capture edge information. The second stage is an adaptive image restoration module, composed of a random noise cancellation submodule and a local contrast enhancement submodule, which collaboratively enhances image details and suppresses noise. In addition, TGAN’s discriminator incorporates multi-scale modules, a strategy that balances global consistency and local detail. Through the interactive game between the generator and discriminator, the generator continuously optimizes the restored image, improving restoration performance. This dynamic adversarial process drives iterative optimization of the generator in the image restoration task. To verify the effectiveness of TGAN, experiments are conducted on the Sen2_MTC dataset. The results show that TGAN significantly outperforms existing methods in terms of peak signal-to-noise ratio (PSNR) and subjective visual evaluation, with PSNR values of 21.547 dB and 20.206 dB for the training and test sets, respectively, indicating that TGAN demonstrates strong performance and application potential in remote sensing cloud image restoration.
Cloud occlusion in optical remote sensing images is one of the core challenges in remote sensing data processing. To address the limitations of current cloud removal technologies in handling cloud edge information and preserving image details, a generative adversarial network (TGAN) based on temporal-spectral domain fusion and temporal self-attention enhancement is proposed. Through its two-stage modular design, TGAN simultaneously improves the quality of remote sensing image restoration and processing efficiency. In the first stage, the feature extraction module, based on a temporal self-attention mechanism, uses a linear expansion layer to capture temporal and spectral domain features, compensating for the limitations of traditional maximum pooling with a one-dimensional linear dimensionality reduction layer, thus enhancing the modeling capability of time-series positional features. A multi-head self-attention mechanism with a weight allocation strategy is designed to accurately capture edge information. The second stage is an adaptive image restoration module, composed of a random noise cancellation submodule and a local contrast enhancement submodule, which collaboratively enhances image details and suppresses noise. In addition, TGAN’s discriminator incorporates multi-scale modules, a strategy that balances global consistency and local detail. Through the interactive game between the generator and discriminator, the generator continuously optimizes the restored image, improving restoration performance. This dynamic adversarial process drives iterative optimization of the generator in the image restoration task. To verify the effectiveness of TGAN, experiments are conducted on the Sen2_MTC dataset. The results show that TGAN significantly outperforms existing methods in terms of peak signal-to-noise ratio (PSNR) and subjective visual evaluation, with PSNR values of 21.547 dB and 20.206 dB for the training and test sets, respectively, indicating that TGAN demonstrates strong performance and application potential in remote sensing cloud image restoration.
Article Search
Search by issue
Select AllDeselectExport
Display Method:
Available online: June 27, 2025
,DOI: 10.15888/j.cnki.csa.009919
Abstract:
Addressing issues such as increased surface cover complexity, heightened heterogeneity within homogeneous regions, and greater similarity between different regions in high-resolution optical remote sensing images, which increase classification difficulty, a supervised learning method based on a dual neighborhood relationships Gaussian regression mixture model (GRMM), an improved version of the Gaussian mixture model (GMM), is proposed. First, supervised sampling of image regions is conducted, with histograms fitted using the least squares method to establish Gaussian mixture models for each land cover, rep-resenting the complex gray-scale features of land cover. Second, local spatial information of adjacent pixels is in-corporated into the image gray-scale space to construct a Gaussian regression model. Finally, in the membership space, neighborhood relationships are processed again to make classification decisions. GRMM achieves kappa coefficients of 97.2% on synthetic images and 98.5% on real high-resolution remote sensing images. Compared to existing mainstream models, GRMM demonstrates strong classification efficiency, noise reduction capability, and generalization ability, with clear classification boundaries, effectively enhancing the classification performance of high-resolution remote sensing images.
Addressing issues such as increased surface cover complexity, heightened heterogeneity within homogeneous regions, and greater similarity between different regions in high-resolution optical remote sensing images, which increase classification difficulty, a supervised learning method based on a dual neighborhood relationships Gaussian regression mixture model (GRMM), an improved version of the Gaussian mixture model (GMM), is proposed. First, supervised sampling of image regions is conducted, with histograms fitted using the least squares method to establish Gaussian mixture models for each land cover, rep-resenting the complex gray-scale features of land cover. Second, local spatial information of adjacent pixels is in-corporated into the image gray-scale space to construct a Gaussian regression model. Finally, in the membership space, neighborhood relationships are processed again to make classification decisions. GRMM achieves kappa coefficients of 97.2% on synthetic images and 98.5% on real high-resolution remote sensing images. Compared to existing mainstream models, GRMM demonstrates strong classification efficiency, noise reduction capability, and generalization ability, with clear classification boundaries, effectively enhancing the classification performance of high-resolution remote sensing images.
Available online: June 27, 2025
,DOI: 10.15888/j.cnki.csa.009923
Abstract:
Aspect-level multimodal sentiment analysis involves analyzing sentiment and opinions at the aspect or attribute level. Previous studies on image-text multimodal sentiment analysis have explored various methods for extracting and fusing features from images and text. Since the initial semantic spaces of images and text are not aligned, common approaches extract deep features from each modality, map them to a shared deep semantic space, and then apply a fusion module. However, this inevitably increases model complexity. With advancements in large language models, aligning the semantic spaces of images and text at a shallow level has become feasible. This study leverages Qwen to generate textual descriptions of images through prompt engineering during preprocessing, transforming multimodal sentiment analysis into a unimodal sentiment analysis task. This enables sentiment analysis results to be obtained using only a text processing module. Experimental results show that, compared to most previous models, the proposed method significantly reduces the number of parameters while achieving performance improvements. Compared to the similarly lightweight TISRI model, it also demonstrates notable advantages in training speed and resource utilization. The code is available at: https://github.com/triangleXIV/ITFFT.
Aspect-level multimodal sentiment analysis involves analyzing sentiment and opinions at the aspect or attribute level. Previous studies on image-text multimodal sentiment analysis have explored various methods for extracting and fusing features from images and text. Since the initial semantic spaces of images and text are not aligned, common approaches extract deep features from each modality, map them to a shared deep semantic space, and then apply a fusion module. However, this inevitably increases model complexity. With advancements in large language models, aligning the semantic spaces of images and text at a shallow level has become feasible. This study leverages Qwen to generate textual descriptions of images through prompt engineering during preprocessing, transforming multimodal sentiment analysis into a unimodal sentiment analysis task. This enables sentiment analysis results to be obtained using only a text processing module. Experimental results show that, compared to most previous models, the proposed method significantly reduces the number of parameters while achieving performance improvements. Compared to the similarly lightweight TISRI model, it also demonstrates notable advantages in training speed and resource utilization. The code is available at: https://github.com/triangleXIV/ITFFT.
Available online: June 27, 2025
,DOI: 10.15888/j.cnki.csa.009921
Abstract:
Federated learning is widely applied in industrial control system intrusion detection, significantly enhancing detection capabilities by integrating high-quality datasets from various systems and collaboratively training high-performance neural network models. However, existing federated learning methods for industrial control systems struggle to balance high performance, low resource consumption, and robust privacy protection. To address this, a dual-server secure serial federated learning (DS-SSFL) approach is proposed for industrial control intrusion detection. The dual-center server coordinates clients for asynchronous serial training, efficiently mining the data features of each client and collaboratively constructing a high-performance intrusion detection model. The approach integrates a decision-differential privacy-preserving mechanism to comprehensively safeguard data privacy and security, while the anti-forgetting aggregation strategy effectively mitigates the catastrophic forgetting issue in serial training. Experimental results show that, compared to traditional federated learning methods, DS-SSFL significantly reduces communication and computational resource overhead, and enhances the model’s robustness and convergence efficiency.
Federated learning is widely applied in industrial control system intrusion detection, significantly enhancing detection capabilities by integrating high-quality datasets from various systems and collaboratively training high-performance neural network models. However, existing federated learning methods for industrial control systems struggle to balance high performance, low resource consumption, and robust privacy protection. To address this, a dual-server secure serial federated learning (DS-SSFL) approach is proposed for industrial control intrusion detection. The dual-center server coordinates clients for asynchronous serial training, efficiently mining the data features of each client and collaboratively constructing a high-performance intrusion detection model. The approach integrates a decision-differential privacy-preserving mechanism to comprehensively safeguard data privacy and security, while the anti-forgetting aggregation strategy effectively mitigates the catastrophic forgetting issue in serial training. Experimental results show that, compared to traditional federated learning methods, DS-SSFL significantly reduces communication and computational resource overhead, and enhances the model’s robustness and convergence efficiency.
Available online: June 27, 2025
,DOI: 10.15888/j.cnki.csa.009950
Abstract:
Visual relocalization remains a widely studied problem in the field of 3D vision, where the goal is to estimate the six degrees of freedom (6DOF) camera pose of a query image within a given prior map. The task is particularly crucial in large-scale indoor environments for applications such as augmented reality and robotic navigation. However, rapid changes in scene appearance, caused by camera movement, pose significant challenges to relocalization systems. To address this issue, a method based on virtual view synthesis is proposed, which enriches the query database and enhances pose estimation in such scenarios. In contrast to approaches that rely on rendering realistic images, the proposed method eliminates the need for high-quality 3D models. Instead, global and local features are directly rendered from virtual viewpoints and are then used for image retrieval and feature matching. Experimental results demonstrate that the proposed method significantly enhances relocalization performance in large-scale indoor settings, achieving improvements of 7.1% and 12.2% on the InLoc dataset.
Visual relocalization remains a widely studied problem in the field of 3D vision, where the goal is to estimate the six degrees of freedom (6DOF) camera pose of a query image within a given prior map. The task is particularly crucial in large-scale indoor environments for applications such as augmented reality and robotic navigation. However, rapid changes in scene appearance, caused by camera movement, pose significant challenges to relocalization systems. To address this issue, a method based on virtual view synthesis is proposed, which enriches the query database and enhances pose estimation in such scenarios. In contrast to approaches that rely on rendering realistic images, the proposed method eliminates the need for high-quality 3D models. Instead, global and local features are directly rendered from virtual viewpoints and are then used for image retrieval and feature matching. Experimental results demonstrate that the proposed method significantly enhances relocalization performance in large-scale indoor settings, achieving improvements of 7.1% and 12.2% on the InLoc dataset.
Available online: June 27, 2025
,DOI: 10.15888/j.cnki.csa.009915
Abstract:
To address challenges such as large-scale variations in target segmentation regions, mis-segmentation of lesion areas, and blurred boundaries in skin images, this study proposes a novel method for skin lesion segmentation, named MSANet. This approach utilizes the pyramid vision Transformer v2 (PVT v2) as the backbone network, integrating the strengths of both Transformer and convolutional neural networks (CNNs). By improving the multi-layer fusion decoding strategy, the proposed method significantly enhances the accuracy of skin lesion segmentation. The decoding process incorporates a split gated attention block (SGA) to capture multi-scale global and local spatial features, thus enhancing the model’s ability to capture contextual information. The multi-scale contextual attention (MCA) module is employed to extract and integrate both channel and positional information, improving the network’s precision in lesion localization. Experimental results on the ISIC2017 and ISIC2018 datasets demonstrate that MSANet achieves Dice scores of 90.12% and 90.91%, and mIoU scores of 85.82% and 84.27%, respectively, outperforming existing methods in segmentation performance.
To address challenges such as large-scale variations in target segmentation regions, mis-segmentation of lesion areas, and blurred boundaries in skin images, this study proposes a novel method for skin lesion segmentation, named MSANet. This approach utilizes the pyramid vision Transformer v2 (PVT v2) as the backbone network, integrating the strengths of both Transformer and convolutional neural networks (CNNs). By improving the multi-layer fusion decoding strategy, the proposed method significantly enhances the accuracy of skin lesion segmentation. The decoding process incorporates a split gated attention block (SGA) to capture multi-scale global and local spatial features, thus enhancing the model’s ability to capture contextual information. The multi-scale contextual attention (MCA) module is employed to extract and integrate both channel and positional information, improving the network’s precision in lesion localization. Experimental results on the ISIC2017 and ISIC2018 datasets demonstrate that MSANet achieves Dice scores of 90.12% and 90.91%, and mIoU scores of 85.82% and 84.27%, respectively, outperforming existing methods in segmentation performance.
Available online: June 27, 2025
,DOI: 10.15888/j.cnki.csa.009912
Abstract:
Infrared small target detection aims to achieve pixel-level separation of small targets from the background in infrared images, with significant applications in military, security, and aerospace fields. However, due to low contrast and low signal-to-noise ratio, existing methods often lose edge information of infrared small targets and fail to effectively utilize the relationship between low-level and high-level features in infrared images. To address these limitations, this study proposes an edge-guided and cross-fusion method for infrared small target detection. Specifically, to overcome the shortcomings of existing methods in extracting edge information, this study constructs an edge-guided feature extraction module. This module integrates edge information into the global-local and detail features of the image through attention weighting, thereby utilizing edge information of small targets more effectively. Additionally, to better fuse high-level and low-level features of the image and enhance the target-background separation capability, this study designs a dual-branch cross-fusion module. This module processes low-level and high-level features of the image through spatial attention and channel attention, respectively, and fully utilizes the complementary relationships between different levels of features through cross-fusion. The experimental results on two benchmark datasets show that compared with state-of-the-art methods, this method improves the IoU metric by 1.89% and the nIoU metric by 2.28%.
Infrared small target detection aims to achieve pixel-level separation of small targets from the background in infrared images, with significant applications in military, security, and aerospace fields. However, due to low contrast and low signal-to-noise ratio, existing methods often lose edge information of infrared small targets and fail to effectively utilize the relationship between low-level and high-level features in infrared images. To address these limitations, this study proposes an edge-guided and cross-fusion method for infrared small target detection. Specifically, to overcome the shortcomings of existing methods in extracting edge information, this study constructs an edge-guided feature extraction module. This module integrates edge information into the global-local and detail features of the image through attention weighting, thereby utilizing edge information of small targets more effectively. Additionally, to better fuse high-level and low-level features of the image and enhance the target-background separation capability, this study designs a dual-branch cross-fusion module. This module processes low-level and high-level features of the image through spatial attention and channel attention, respectively, and fully utilizes the complementary relationships between different levels of features through cross-fusion. The experimental results on two benchmark datasets show that compared with state-of-the-art methods, this method improves the IoU metric by 1.89% and the nIoU metric by 2.28%.
Available online: June 27, 2025
,DOI: 10.15888/j.cnki.csa.009896
Abstract:
Using YOLOv5s as the object detector in DeepSORT presents challenges, including high computational costs, complex in structure, and exhibits limitations in detection accuracy. Firstly, the GhostNet lightweight module is introduced to lightweight the YOLOv5s model, reducing both the number of parameters and computational load of the model to meet the deployment requirements for mobile devices. Secondly, the ECA attention mechanism is incorporated to enhance the model’s perceptual capability, improving detection performance. Lastly, knowledge distillation is applied to the YOLOv5s model to further enhance detection accuracy. The improved YOLOv5s algorithm shows a 2% increase in precision, 1% increase in recall, and 0.8% increase in mAP@0.5, compared to the original algorithm. The model’s parameter count is reduced by 47%, and its complexity by 48%. When the enhanced YOLOv5s is combined with the DeepSORT algorithm, MOTA, MOTP, and IDF1 are improved by 1.2%, 3.1%, and 2.7%, respectively, while IDS is reduced by 35%, compared to the original version. Experimental results verify that the improved YOLOv5s, as a detector, enhances detection speed, reduces pedestrian ID switching, and can be effectively applied to pedestrian tracking.
Using YOLOv5s as the object detector in DeepSORT presents challenges, including high computational costs, complex in structure, and exhibits limitations in detection accuracy. Firstly, the GhostNet lightweight module is introduced to lightweight the YOLOv5s model, reducing both the number of parameters and computational load of the model to meet the deployment requirements for mobile devices. Secondly, the ECA attention mechanism is incorporated to enhance the model’s perceptual capability, improving detection performance. Lastly, knowledge distillation is applied to the YOLOv5s model to further enhance detection accuracy. The improved YOLOv5s algorithm shows a 2% increase in precision, 1% increase in recall, and 0.8% increase in mAP@0.5, compared to the original algorithm. The model’s parameter count is reduced by 47%, and its complexity by 48%. When the enhanced YOLOv5s is combined with the DeepSORT algorithm, MOTA, MOTP, and IDF1 are improved by 1.2%, 3.1%, and 2.7%, respectively, while IDS is reduced by 35%, compared to the original version. Experimental results verify that the improved YOLOv5s, as a detector, enhances detection speed, reduces pedestrian ID switching, and can be effectively applied to pedestrian tracking.
Available online: June 24, 2025
,DOI: 10.15888/j.cnki.csa.009893
Abstract:
High-throughput X-ray diffraction (XRD) analysis plays a crucial role in accelerating material discovery. However, traditional methods often rely heavily on manual interpretation and tend to overlook low-intensity peak information when processing complex XRD data, thus limiting the potential for accuracy improvement. To address this issue, a large language model (LLM) agent framework for multi-modal crystal structure prediction is proposed. The framework integrates a GPT-4-driven intelligent agent with multi-modal voting models based on XRD and pair distribution functions, enabling autonomous crystal structure and space group prediction tasks. In addition, the reasoning capability of the LLM is enhanced through the introduction of knowledge graphs, which aid in understanding the relationships between crystal features, thus improving both prediction accuracy and reasoning performance. Experimental results demonstrate that the accuracy of this framework in crystal structure prediction and space group prediction tasks reaches 97.5% and 98.7%, respectively. This design significantly enhances the accuracy and efficiency of high-throughput analysis, with the potential to advance materials science research and provide valuable insights for addressing other highly interrelated multi-task problems.
High-throughput X-ray diffraction (XRD) analysis plays a crucial role in accelerating material discovery. However, traditional methods often rely heavily on manual interpretation and tend to overlook low-intensity peak information when processing complex XRD data, thus limiting the potential for accuracy improvement. To address this issue, a large language model (LLM) agent framework for multi-modal crystal structure prediction is proposed. The framework integrates a GPT-4-driven intelligent agent with multi-modal voting models based on XRD and pair distribution functions, enabling autonomous crystal structure and space group prediction tasks. In addition, the reasoning capability of the LLM is enhanced through the introduction of knowledge graphs, which aid in understanding the relationships between crystal features, thus improving both prediction accuracy and reasoning performance. Experimental results demonstrate that the accuracy of this framework in crystal structure prediction and space group prediction tasks reaches 97.5% and 98.7%, respectively. This design significantly enhances the accuracy and efficiency of high-throughput analysis, with the potential to advance materials science research and provide valuable insights for addressing other highly interrelated multi-task problems.
Available online: June 24, 2025
,DOI: 10.15888/j.cnki.csa.009909
Abstract:
In recent years, aspect-based sentiment analysis utilizing graph neural networks to exploit syntactic information has become increasingly popular. However, existing methods often overlook the impact of different types of relations on content words, making it challenging to distinguish key relational terms. In addition, the mutual supplementation of multi-perspective information plays a crucial role in capturing sentiment features, but integration mechanisms have often been neglected in past research. To address these issues, a multi-source information graph convolutional network (MSI-GCN) is proposed to effectively capture and integrate three perspectives of information. First, a dual-channel information extraction module, syntax-semantics dual graph convolutional network (SSD-GCN), is designed, comprising a type-enhanced syntax graph convolutional network (TES-GCN) and a semantic graph convolutional network (SEM-GCN). TES-GCN enhances syntactic information by incorporating a type embedding layer and learning different weights through the syntactic module. SEM-GCN encodes self-attention matrices to capture semantic information and employs orthogonal regularization to strengthen semantic associations. Second, external knowledge graph embeddings enrich vocabulary features. Finally, a local-global convolutional network is introduced to leverage the complementarity between perspectives for effective integration. The proposed method is evaluated on four public datasets, showing improvements in accuracy and Macro-F1 scores compared to baseline models.
In recent years, aspect-based sentiment analysis utilizing graph neural networks to exploit syntactic information has become increasingly popular. However, existing methods often overlook the impact of different types of relations on content words, making it challenging to distinguish key relational terms. In addition, the mutual supplementation of multi-perspective information plays a crucial role in capturing sentiment features, but integration mechanisms have often been neglected in past research. To address these issues, a multi-source information graph convolutional network (MSI-GCN) is proposed to effectively capture and integrate three perspectives of information. First, a dual-channel information extraction module, syntax-semantics dual graph convolutional network (SSD-GCN), is designed, comprising a type-enhanced syntax graph convolutional network (TES-GCN) and a semantic graph convolutional network (SEM-GCN). TES-GCN enhances syntactic information by incorporating a type embedding layer and learning different weights through the syntactic module. SEM-GCN encodes self-attention matrices to capture semantic information and employs orthogonal regularization to strengthen semantic associations. Second, external knowledge graph embeddings enrich vocabulary features. Finally, a local-global convolutional network is introduced to leverage the complementarity between perspectives for effective integration. The proposed method is evaluated on four public datasets, showing improvements in accuracy and Macro-F1 scores compared to baseline models.
Available online: June 24, 2025
,DOI: 10.15888/j.cnki.csa.009910
Abstract:
In recent years, with the increasing demand for diagnosing autism spectrum disorder (ASD), automated detection methods based on electroencephalogram (EEG) signals have gained significant attention. However, existing approaches still face challenges in terms of accuracy, generalization ability, robustness, and interpretability. An improved ASD detection model, attention-CosCNN-LSTM-net (ACLNet), is proposed. A multi-dimensional attention mechanism is leveraged to enhance the focus on critical signals. A cosine convolutional neural network is combined to capture frequency features of EEG signals, and a tree-structured LSTM module is integrated to model the hierarchical structure and long-term dependencies within signals. These components enable comprehensive extraction of spatiotemporal and frequency domain features of EEG signals. Experiments conducted on an ASD EEG dataset with five-fold cross-validation demonstrate that ACLNet achieves a classification accuracy of 94.11%, a recall rate of 93.29%, and a precision rate of 93.78%, significantly outperforming existing detection methods. Moreover, the model exhibits stable performance across different data splits and unseen data, validating strong generalization ability and robustness. Ablation studies further confirm the critical contributions of each module to feature extraction and overall performance. This study provides an efficient, stable, and interpretable solution for automated ASD detection, advancing the application of EEG signals in ASD diagnosis and offering valuable support for related research and clinical practice.
In recent years, with the increasing demand for diagnosing autism spectrum disorder (ASD), automated detection methods based on electroencephalogram (EEG) signals have gained significant attention. However, existing approaches still face challenges in terms of accuracy, generalization ability, robustness, and interpretability. An improved ASD detection model, attention-CosCNN-LSTM-net (ACLNet), is proposed. A multi-dimensional attention mechanism is leveraged to enhance the focus on critical signals. A cosine convolutional neural network is combined to capture frequency features of EEG signals, and a tree-structured LSTM module is integrated to model the hierarchical structure and long-term dependencies within signals. These components enable comprehensive extraction of spatiotemporal and frequency domain features of EEG signals. Experiments conducted on an ASD EEG dataset with five-fold cross-validation demonstrate that ACLNet achieves a classification accuracy of 94.11%, a recall rate of 93.29%, and a precision rate of 93.78%, significantly outperforming existing detection methods. Moreover, the model exhibits stable performance across different data splits and unseen data, validating strong generalization ability and robustness. Ablation studies further confirm the critical contributions of each module to feature extraction and overall performance. This study provides an efficient, stable, and interpretable solution for automated ASD detection, advancing the application of EEG signals in ASD diagnosis and offering valuable support for related research and clinical practice.
Available online: June 24, 2025
,DOI: 10.15888/j.cnki.csa.009911
Abstract:
In the field of process mining, numerous process operations are highly dependent on accurate timestamp information in event logs. Consequently, quality issues related to timestamps have particularly significant impacts, especially identical timestamp errors, which can lead to misleading process insights and subsequently result in severe process deviations. Existing research addressing such errors lacks sufficient consideration of long-term dependencies between events and potential correlations between attributes, limiting the accuracy of repairing identical timestamp errors. To address this issue, a method for repairing identical timestamp errors based on a hierarchical Transformer model is proposed. This method captures long-distance behavioral dependencies between events and deep correlation information between attributes through hierarchical information propagation combined with multi-view interaction. The tasks of reordering erroneous events and predicting corresponding timestamps are accomplished layer by layer, thus effectively repairing event logs with identical timestamp errors. Evaluations conducted on four publicly available datasets demonstrate that the proposed method can effectively improve the accuracy of repairing identical timestamp errors.
In the field of process mining, numerous process operations are highly dependent on accurate timestamp information in event logs. Consequently, quality issues related to timestamps have particularly significant impacts, especially identical timestamp errors, which can lead to misleading process insights and subsequently result in severe process deviations. Existing research addressing such errors lacks sufficient consideration of long-term dependencies between events and potential correlations between attributes, limiting the accuracy of repairing identical timestamp errors. To address this issue, a method for repairing identical timestamp errors based on a hierarchical Transformer model is proposed. This method captures long-distance behavioral dependencies between events and deep correlation information between attributes through hierarchical information propagation combined with multi-view interaction. The tasks of reordering erroneous events and predicting corresponding timestamps are accomplished layer by layer, thus effectively repairing event logs with identical timestamp errors. Evaluations conducted on four publicly available datasets demonstrate that the proposed method can effectively improve the accuracy of repairing identical timestamp errors.
Available online: June 24, 2025
,DOI: 10.15888/j.cnki.csa.009914
Abstract:
Buffer overflow vulnerabilities are widely present in programs written in insecure high-level languages. By exploiting buffer overflow vulnerabilities, attackers can carry out dangerous attacks such as control flow hijacking. Canary-based stack smashing protection is a simple, effective, and widely deployed defense mechanism against buffer overflow vulnerabilities. However, its characteristics of fixed locations and identical values make it susceptible to analysis and exploitation by attackers. This study proposes a stack smashing protection approach based on software diversity, which features heterogeneous canary with randomized sizes and offsets. This approach not only directly defends against leakage and overwrite attacks that conventional canary cannot handle, but also enables the construction of various diversified software systems with more security. Experimental results show that this technology introduces less than 2% compilation overhead and an average of 3.22% runtime overhead for the SPEC CPU 2017 benchmark set while effectively improving security.
Buffer overflow vulnerabilities are widely present in programs written in insecure high-level languages. By exploiting buffer overflow vulnerabilities, attackers can carry out dangerous attacks such as control flow hijacking. Canary-based stack smashing protection is a simple, effective, and widely deployed defense mechanism against buffer overflow vulnerabilities. However, its characteristics of fixed locations and identical values make it susceptible to analysis and exploitation by attackers. This study proposes a stack smashing protection approach based on software diversity, which features heterogeneous canary with randomized sizes and offsets. This approach not only directly defends against leakage and overwrite attacks that conventional canary cannot handle, but also enables the construction of various diversified software systems with more security. Experimental results show that this technology introduces less than 2% compilation overhead and an average of 3.22% runtime overhead for the SPEC CPU 2017 benchmark set while effectively improving security.
Available online: June 24, 2025
,DOI: 10.15888/j.cnki.csa.009949
Abstract:
In federated learning, unstable clients may disrupt the training process of the global model through data pollution or malicious behavior. Traditional defense methods typically focus on excluding these clients, but overlook the fact that the data generated by unstable clients can also provide valuable training signals for the model. To address this issue, a federated adversarial training with an enhanced adaptability method, Fed-ATEA, which utilizes adversarial examples generated by unstable clients to enhance the robustness of the global model is proposed. This framework allows for the incorporation of adversarial samples into the training process of trusted clients without excluding unstable ones, thereby improving the stability and robustness of the model. By dynamically adjusting training strategies, Fed-ATEA maximizes the utilization of beneficial signals from unstable clients while effectively mitigating their negative influences. Experimental results show that, compared to other federated learning methods, Fed-ATEA demonstrates stronger stability and robustness in handling attacks and noise interference.
In federated learning, unstable clients may disrupt the training process of the global model through data pollution or malicious behavior. Traditional defense methods typically focus on excluding these clients, but overlook the fact that the data generated by unstable clients can also provide valuable training signals for the model. To address this issue, a federated adversarial training with an enhanced adaptability method, Fed-ATEA, which utilizes adversarial examples generated by unstable clients to enhance the robustness of the global model is proposed. This framework allows for the incorporation of adversarial samples into the training process of trusted clients without excluding unstable ones, thereby improving the stability and robustness of the model. By dynamically adjusting training strategies, Fed-ATEA maximizes the utilization of beneficial signals from unstable clients while effectively mitigating their negative influences. Experimental results show that, compared to other federated learning methods, Fed-ATEA demonstrates stronger stability and robustness in handling attacks and noise interference.
Available online: June 24, 2025
,DOI: 10.15888/j.cnki.csa.009944
Abstract:
A fall detection algorithm combining OpenPose with an improved ST-GCN is proposed to address the limitations of low accuracy and the need for pre-defining human skeleton topology graphs of the ST-GCN algorithm in action recognition. The OpenPose algorithm is used to extract the human skeletal keypoint data, which is then input into the improved ST-GCN algorithm for action recognition. The ST-GCN algorithm is improved by introducing an adaptive graph convolution module, which dynamically adjusts the graph structure to enhance the flexibility in feature extraction across different action types; an attention mechanism module is introduced to further improve the recognition performance of the model. Validation on publicly available datasets shows that on the NTU-RGB+D 60 dataset, the top-1 accuracy of X-Sub and X-View is improved by 2.2% and 2.5%, respectively, compared with the baseline; on the Kinetics-Skeleton dataset, the top-1 and top-5 accuracy are improved by 3.1% and 4%, respectively. In addition, the accuracy measured on the self-constructed dataset is improved by 4.7% compared with that before the improvement. The experimental results show that the proposed algorithm meets the requirements of practical applications.
A fall detection algorithm combining OpenPose with an improved ST-GCN is proposed to address the limitations of low accuracy and the need for pre-defining human skeleton topology graphs of the ST-GCN algorithm in action recognition. The OpenPose algorithm is used to extract the human skeletal keypoint data, which is then input into the improved ST-GCN algorithm for action recognition. The ST-GCN algorithm is improved by introducing an adaptive graph convolution module, which dynamically adjusts the graph structure to enhance the flexibility in feature extraction across different action types; an attention mechanism module is introduced to further improve the recognition performance of the model. Validation on publicly available datasets shows that on the NTU-RGB+D 60 dataset, the top-1 accuracy of X-Sub and X-View is improved by 2.2% and 2.5%, respectively, compared with the baseline; on the Kinetics-Skeleton dataset, the top-1 and top-5 accuracy are improved by 3.1% and 4%, respectively. In addition, the accuracy measured on the self-constructed dataset is improved by 4.7% compared with that before the improvement. The experimental results show that the proposed algorithm meets the requirements of practical applications.
Available online: June 24, 2025
,DOI: 10.15888/j.cnki.csa.009924
Abstract:
Addressing the challenge of efficiently fusing audio and video features while accurately extracting time-dependent emotion information in audiovisual emotion recognition, a mutual information-based audiovisual emotion recognition model is proposed, incorporating Kolmogorov-Arnold long short-term memory (KLSTM). Feature selection and adaptive window processing, based on the mutual information approach, are employed to extract emotionally relevant key segments from audio and video signals, effectively reducing information redundancy. The KLSTM network is integrated into feature extraction to capture the temporal dependencies of audiovisual modal signals. In the fusion stage, cross-modal consistency maximization ensures the coordination and complementarity of audio and video features. Experimental results demonstrate that the proposed model outperforms existing benchmark models on both CMU-MOSI and CMU-MOSEI datasets, validating its effectiveness in multimodal emotion recognition tasks.
Addressing the challenge of efficiently fusing audio and video features while accurately extracting time-dependent emotion information in audiovisual emotion recognition, a mutual information-based audiovisual emotion recognition model is proposed, incorporating Kolmogorov-Arnold long short-term memory (KLSTM). Feature selection and adaptive window processing, based on the mutual information approach, are employed to extract emotionally relevant key segments from audio and video signals, effectively reducing information redundancy. The KLSTM network is integrated into feature extraction to capture the temporal dependencies of audiovisual modal signals. In the fusion stage, cross-modal consistency maximization ensures the coordination and complementarity of audio and video features. Experimental results demonstrate that the proposed model outperforms existing benchmark models on both CMU-MOSI and CMU-MOSEI datasets, validating its effectiveness in multimodal emotion recognition tasks.
Available online: June 20, 2025
,DOI: 10.15888/j.cnki.csa.009941
Abstract:
Visual-inertial odometry (VIO) achieves pose estimation by fusing visual and inertial data. In complex environments, inertial data are prone to noise interference, and long-term motion leads to cumulative errors. Additionally, most VIO models overlook local information interaction between modalities and fail to fully utilize their complementary nature, thereby compromising pose estimation accuracy. To address these issues, this study proposes an attention and local interaction-based visual-inertial odometry (ALVIO) model. First, the model extracts visual features and inertial features. Then, the historical time-series information of the inertial features is preserved, and a channel attention mechanism based on discrete cosine transform (DCT) is applied to enhance low-frequency effective features and suppress high-frequency noise. Next, a multi-modal local interaction and global fusion module is designed, which gradually achieves local interaction and global fusion between modalities through improved split-attention mechanism and MLP-Mixer. This module adjusts the local feature weights based on the contributions of different modalities to realize inter-modal complementarity and then integrates the features globally to obtain a unified representation. Finally, the fused features are used for temporal modeling and pose regression to obtain relative poses. To verify the effectiveness of the model in complex environments, this paper conducts experiments on processed low-quality versions of the public KITTI and EuRoC datasets. The results show that, compared to the direct feature concatenation model, the multi-head attention fusion model, and the soft mask fusion model, ALVIO reduces the translation error by 49.92%, 32.82%, and 37.74%, respectively, and the rotation error by 51.34%, 25.96%, and 29.54%, respectively, while also demonstrating higher efficiency and robustness.
Visual-inertial odometry (VIO) achieves pose estimation by fusing visual and inertial data. In complex environments, inertial data are prone to noise interference, and long-term motion leads to cumulative errors. Additionally, most VIO models overlook local information interaction between modalities and fail to fully utilize their complementary nature, thereby compromising pose estimation accuracy. To address these issues, this study proposes an attention and local interaction-based visual-inertial odometry (ALVIO) model. First, the model extracts visual features and inertial features. Then, the historical time-series information of the inertial features is preserved, and a channel attention mechanism based on discrete cosine transform (DCT) is applied to enhance low-frequency effective features and suppress high-frequency noise. Next, a multi-modal local interaction and global fusion module is designed, which gradually achieves local interaction and global fusion between modalities through improved split-attention mechanism and MLP-Mixer. This module adjusts the local feature weights based on the contributions of different modalities to realize inter-modal complementarity and then integrates the features globally to obtain a unified representation. Finally, the fused features are used for temporal modeling and pose regression to obtain relative poses. To verify the effectiveness of the model in complex environments, this paper conducts experiments on processed low-quality versions of the public KITTI and EuRoC datasets. The results show that, compared to the direct feature concatenation model, the multi-head attention fusion model, and the soft mask fusion model, ALVIO reduces the translation error by 49.92%, 32.82%, and 37.74%, respectively, and the rotation error by 51.34%, 25.96%, and 29.54%, respectively, while also demonstrating higher efficiency and robustness.
Available online: June 20, 2025
,DOI: 10.15888/j.cnki.csa.009922
Abstract:
Dynamic binary translation (DBT) is an efficient technology for instruction set simulation, commonly used to build CPU simulation models. However, significant challenges arise when simulating digital signal processors (DSP). High-performance DSP, such as TI’s TMS320C6X series, often adopt very long instruction word (VLIW) architectures and include specialized hardware and instructions to simplify the use of software-pipelined loops. The absence of explicit conditional branch instructions and loop counter modification operations in such loops, along with the reordering, overlapping, and masking of instructions within the loops, makes translating these loops using DBT technology highly challenging. To address this, this study proposes a novel dynamic translation approach. The approach involves serializing parallel loop iterations, generating translation blocks for different states, and overlapping the instructions of inner and outer loops while aligning them by cycle to synchronize the translation of outer and inner loops, thereby accurately simulating the execution of software-pipelined loops. Experimental results show that when running commonly used code containing software pipelining (e.g. dsplib), the proposed simulator employing this translation approach produces results identical to those from a hardware development board, demonstrating the correctness of the solution. Moreover, the proposed simulator performs 3.25× faster than TI’s official simulator.
Dynamic binary translation (DBT) is an efficient technology for instruction set simulation, commonly used to build CPU simulation models. However, significant challenges arise when simulating digital signal processors (DSP). High-performance DSP, such as TI’s TMS320C6X series, often adopt very long instruction word (VLIW) architectures and include specialized hardware and instructions to simplify the use of software-pipelined loops. The absence of explicit conditional branch instructions and loop counter modification operations in such loops, along with the reordering, overlapping, and masking of instructions within the loops, makes translating these loops using DBT technology highly challenging. To address this, this study proposes a novel dynamic translation approach. The approach involves serializing parallel loop iterations, generating translation blocks for different states, and overlapping the instructions of inner and outer loops while aligning them by cycle to synchronize the translation of outer and inner loops, thereby accurately simulating the execution of software-pipelined loops. Experimental results show that when running commonly used code containing software pipelining (e.g. dsplib), the proposed simulator employing this translation approach produces results identical to those from a hardware development board, demonstrating the correctness of the solution. Moreover, the proposed simulator performs 3.25× faster than TI’s official simulator.
Available online: June 20, 2025
,DOI: 10.15888/j.cnki.csa.009951
Abstract:
With the significant progress of satellite video imaging technology, object tracking in satellite videos has attracted more and more researchers’ attention. However, most of the previous research obtains spatial information through the global attention mechanism, which makes the model focus on the background part and thus ignore the object; moreover, only spatial information of the object in the video frames is utilized, resulting in inaccurate object localization. In this study, we improve the existing Siamese network object tracking model SiamCAR and a spatio-temporal Siamese network Siam-STM. Specifically, we proposes a spatial information perception module based on the attention mechanism, which aggregates the contextual information in the images and enhances the discriminative capability of small object features in the satellite videos; to utilize the temporal information across video frames, a temporal information perception module is proposed to fuse the current frame with the historical frames, enabling the position information of the object across time to be learned, the object’s trajectory to be better tracked, and the interference from similar objects to be mitigated. In addition, to mitigate the effects of occlusion in satellite videos, this study introduces a linear fitting method based on the Kalman filter and then proposes a motion estimation mechanism. This mechanism can effectively model the motion characteristics of the object, allowing accurate localization even during occlusions. The effectiveness of Siam-STM is verified by comparing it with state-of-the-art models on the SatSOT dataset.
With the significant progress of satellite video imaging technology, object tracking in satellite videos has attracted more and more researchers’ attention. However, most of the previous research obtains spatial information through the global attention mechanism, which makes the model focus on the background part and thus ignore the object; moreover, only spatial information of the object in the video frames is utilized, resulting in inaccurate object localization. In this study, we improve the existing Siamese network object tracking model SiamCAR and a spatio-temporal Siamese network Siam-STM. Specifically, we proposes a spatial information perception module based on the attention mechanism, which aggregates the contextual information in the images and enhances the discriminative capability of small object features in the satellite videos; to utilize the temporal information across video frames, a temporal information perception module is proposed to fuse the current frame with the historical frames, enabling the position information of the object across time to be learned, the object’s trajectory to be better tracked, and the interference from similar objects to be mitigated. In addition, to mitigate the effects of occlusion in satellite videos, this study introduces a linear fitting method based on the Kalman filter and then proposes a motion estimation mechanism. This mechanism can effectively model the motion characteristics of the object, allowing accurate localization even during occlusions. The effectiveness of Siam-STM is verified by comparing it with state-of-the-art models on the SatSOT dataset.
ZHANG Jiu-Ling, ZHU Mei-Ling, XU Xin-Run, WU Yu-Rong, YANG Jian-Wen, ZHANG Qiu-Hong, WANG Wei, WANG Jia, JIANG Hui-Long
Available online: June 20, 2025
,DOI: 10.15888/j.cnki.csa.009907
Abstract:
Monocular depth estimation (MDE) is a core task in computer vision, essential for spatial understanding, 3D reconstruction, autonomous driving, and other applications. Deep learning-based MDE methods can predict relative depth from a single image, but they often lack metric information, leading to scale inconsistency and limiting their utility in downstream tasks such as visual SLAM, 3D reconstruction, and novel view synthesis. To address these limitations, monocular metric depth estimation (MMDE) has been introduced. By enabling scene-scale inference, MMDE addresses depth consistency issues, enhances temporal stability in sequential tasks, and simplifies the integration of downstream applications, significantly expanding its practical use cases. This study offers a comprehensive review of the development of depth estimation technologies, from traditional geometric approaches to modern deep learning methods, highlighting key milestones and breakthroughs in the field. Special emphasis is placed on scale-agnostic methods and their role in enabling zero-shot generalization, which has been foundational for the progress of MMDE. The study also examines the latest advancements in zero-shot MMDE, focusing on critical challenges such as improving model generalization and preserving fine edge details. To address these issues, innovative solutions have been explored, including advanced data augmentation techniques, refined model architectures, and the integration of generative approaches, leading to significant advancements. This review analyzes these solutions and their contributions in depth. The study concludes by synthesizing the connections between recent achievements in zero-shot MMDE, identifying unresolved challenges, and exploring potential future directions for research. By providing an in-depth analysis of the current state of the field and emerging trends, this study aims to serve as a valuable resource for researchers, offering a clear roadmap for understanding and advancing MMDE technology across a wide range of applications.
Monocular depth estimation (MDE) is a core task in computer vision, essential for spatial understanding, 3D reconstruction, autonomous driving, and other applications. Deep learning-based MDE methods can predict relative depth from a single image, but they often lack metric information, leading to scale inconsistency and limiting their utility in downstream tasks such as visual SLAM, 3D reconstruction, and novel view synthesis. To address these limitations, monocular metric depth estimation (MMDE) has been introduced. By enabling scene-scale inference, MMDE addresses depth consistency issues, enhances temporal stability in sequential tasks, and simplifies the integration of downstream applications, significantly expanding its practical use cases. This study offers a comprehensive review of the development of depth estimation technologies, from traditional geometric approaches to modern deep learning methods, highlighting key milestones and breakthroughs in the field. Special emphasis is placed on scale-agnostic methods and their role in enabling zero-shot generalization, which has been foundational for the progress of MMDE. The study also examines the latest advancements in zero-shot MMDE, focusing on critical challenges such as improving model generalization and preserving fine edge details. To address these issues, innovative solutions have been explored, including advanced data augmentation techniques, refined model architectures, and the integration of generative approaches, leading to significant advancements. This review analyzes these solutions and their contributions in depth. The study concludes by synthesizing the connections between recent achievements in zero-shot MMDE, identifying unresolved challenges, and exploring potential future directions for research. By providing an in-depth analysis of the current state of the field and emerging trends, this study aims to serve as a valuable resource for researchers, offering a clear roadmap for understanding and advancing MMDE technology across a wide range of applications.
Available online: June 20, 2025
,DOI: 10.15888/j.cnki.csa.009945
Abstract:
To address the problem that existing algorithms cannot effectively integrate local details and global structures, this study proposes a three-stage face image restoration algorithm that incorporates attention mechanisms to optimize local and global features. In the first stage, the position attention module (PAM) and focused linear attention (FLA) are introduced to enhance the extraction of local texture details and global contextual features of the image. In the second stage of optimization, the convolutional block attention module (CBAM) is incorporated with skip connections. The design strengthens feature focus through the differentiated weight assignment in both channel and spatial dimensions, while effectively achieving refined reconstruction of local regions by utilizing the detail-preservation strategy during the downsampling process. Finally, the third stage of integrating features is introduced to make the restored image more robust. The experimental results show that the proposed method achieves average improvements of 0.1214 dB in PSNR and 0.0022 in SSIM, along with an average reduction of 0.00065 in LPIPS on the CelebA-HQ dataset, significantly enhancing both the restoration quality and visual perception of images.
To address the problem that existing algorithms cannot effectively integrate local details and global structures, this study proposes a three-stage face image restoration algorithm that incorporates attention mechanisms to optimize local and global features. In the first stage, the position attention module (PAM) and focused linear attention (FLA) are introduced to enhance the extraction of local texture details and global contextual features of the image. In the second stage of optimization, the convolutional block attention module (CBAM) is incorporated with skip connections. The design strengthens feature focus through the differentiated weight assignment in both channel and spatial dimensions, while effectively achieving refined reconstruction of local regions by utilizing the detail-preservation strategy during the downsampling process. Finally, the third stage of integrating features is introduced to make the restored image more robust. The experimental results show that the proposed method achieves average improvements of 0.1214 dB in PSNR and 0.0022 in SSIM, along with an average reduction of 0.00065 in LPIPS on the CelebA-HQ dataset, significantly enhancing both the restoration quality and visual perception of images.
Available online: June 20, 2025
,DOI: 10.15888/j.cnki.csa.009943
Abstract:
To improve the accuracy and robustness of modulation recognition, this study proposes an improved bimodal hybrid modulation recognition model. The model incorporates both the original time-domain in-phase and quadrature (I/Q) data, as well as amplitude and phase (A/P) format data to explore the spatiotemporal correlations within the signal. A two-branch symmetric structure is applied to further process the A/P data, enabling more effective learning of repetitive patterns while mitigating information redundancy. A bidirectional long short-term memory (BiLSTM) network is introduced to enhance the model’s capacity for complex temporal feature extraction. Experimental results demonstrate that the proposed model performs well on the RadioML2016.10A dataset. When the signal-to-noise ratio (SNR) is below –8 dB, the average recognition accuracy surpasses mainstream models by 6%. Within the SNR range of 0 to 18 dB, the average recognition accuracy improves by 2% to 10%, reaching 94.32% at 16 dB. In addition, when applied to the RadioML2016.10B dataset, the model continues to achieve superior performance, attaining a recognition accuracy of 93.91% at 18 dB.
To improve the accuracy and robustness of modulation recognition, this study proposes an improved bimodal hybrid modulation recognition model. The model incorporates both the original time-domain in-phase and quadrature (I/Q) data, as well as amplitude and phase (A/P) format data to explore the spatiotemporal correlations within the signal. A two-branch symmetric structure is applied to further process the A/P data, enabling more effective learning of repetitive patterns while mitigating information redundancy. A bidirectional long short-term memory (BiLSTM) network is introduced to enhance the model’s capacity for complex temporal feature extraction. Experimental results demonstrate that the proposed model performs well on the RadioML2016.10A dataset. When the signal-to-noise ratio (SNR) is below –8 dB, the average recognition accuracy surpasses mainstream models by 6%. Within the SNR range of 0 to 18 dB, the average recognition accuracy improves by 2% to 10%, reaching 94.32% at 16 dB. In addition, when applied to the RadioML2016.10B dataset, the model continues to achieve superior performance, attaining a recognition accuracy of 93.91% at 18 dB.
Available online: June 13, 2025
,DOI: 10.15888/j.cnki.csa.009913
Abstract:
Due to the intermittency, instability, and randomness of solar energy, accurate short-term photovoltaic (PV) power forecasting presents significant challenges, hindering the integration of photovoltaic systems with smart grids. To address this, a method called WOA-VMD and PSO-DSN (WVPD) is proposed. First, variational mode decomposition (VMD) is applied to obtain multiple Intrinsic mode functions (IMFs) components. Meanwhile, the whale optimization algorithm (WOA) is used to optimize the mode components and penalty factor parameters to resolve issues of VMD decomposition and mode mixing. Then, the spatial and temporal correlations of PV power and numerical weather prediction (NWP) data are utilized to construct a novel dual-stream network (DSN), which combines squeeze-and-excitation networks (SENet) and bidirectional gated recurrent units (BiGRU). In addition, particle swarm optimization (PSO) is used to optimize the learning rate and batch size in the DSN. Finally, experiments demonstrate that compared to deep learning hybrid models, the proposed WVPD method improves MSE by 78.6%, RMSE by 53.7%, and MAE by 37.7%, showing superior performance over state-of-the-art models. The code can be found at https://github.com/ruanyuyuan/PV-power-forecast.
Due to the intermittency, instability, and randomness of solar energy, accurate short-term photovoltaic (PV) power forecasting presents significant challenges, hindering the integration of photovoltaic systems with smart grids. To address this, a method called WOA-VMD and PSO-DSN (WVPD) is proposed. First, variational mode decomposition (VMD) is applied to obtain multiple Intrinsic mode functions (IMFs) components. Meanwhile, the whale optimization algorithm (WOA) is used to optimize the mode components and penalty factor parameters to resolve issues of VMD decomposition and mode mixing. Then, the spatial and temporal correlations of PV power and numerical weather prediction (NWP) data are utilized to construct a novel dual-stream network (DSN), which combines squeeze-and-excitation networks (SENet) and bidirectional gated recurrent units (BiGRU). In addition, particle swarm optimization (PSO) is used to optimize the learning rate and batch size in the DSN. Finally, experiments demonstrate that compared to deep learning hybrid models, the proposed WVPD method improves MSE by 78.6%, RMSE by 53.7%, and MAE by 37.7%, showing superior performance over state-of-the-art models. The code can be found at https://github.com/ruanyuyuan/PV-power-forecast.
Available online: June 13, 2025
,DOI: 10.15888/j.cnki.csa.009916
Abstract:
Addressing the challenges in extracting subtle fingerprint features of communication emitters and the low recognition rate of single-feature identification, a method for individual identification of communication emitters is proposed. This method combines variational mode decomposition (VMD) with multi-domain feature fusion (MDFF), based on the joint application of minimum Bhattacharyya distance (BDmin) and correlation analysis (CA). First, the VMD method based on BDmin is utilized to decompose each symbol waveform of the communication emitter signal into several intrinsic mode functions (IMFs), including low-frequency IMFs containing data information and high-frequency IMFs containing fingerprint information. Next, the correlation coefficients between each IMF and its symbol waveform signal are calculated, with IMFs exhibiting small correlation coefficients being selected as the subtle feature components of the emitter. Time-domain, frequency-domain, and entropy-based multi-feature parameters are then extracted from these subtle feature components and concatenated into a multi-domain feature vector for feature extraction of the communication emitter’s symbol waveform. Finally, a long short-term memory (LSTM) network is employed to sequentially learn and classify the multi-domain feature vectors of each symbol of the emitter signal, thus achieving individual identification and classification of communication emitters. Experimental validation is conducted using the publicly available Oracle dataset. The results show that, when the signal-to-noise ratio (SNR) is 6 dB, the proposed algorithm achieves a recognition accuracy of 96.7%, representing a 22.1% improvement compared to the average reco gnition accuracy of individual domains.
Addressing the challenges in extracting subtle fingerprint features of communication emitters and the low recognition rate of single-feature identification, a method for individual identification of communication emitters is proposed. This method combines variational mode decomposition (VMD) with multi-domain feature fusion (MDFF), based on the joint application of minimum Bhattacharyya distance (BDmin) and correlation analysis (CA). First, the VMD method based on BDmin is utilized to decompose each symbol waveform of the communication emitter signal into several intrinsic mode functions (IMFs), including low-frequency IMFs containing data information and high-frequency IMFs containing fingerprint information. Next, the correlation coefficients between each IMF and its symbol waveform signal are calculated, with IMFs exhibiting small correlation coefficients being selected as the subtle feature components of the emitter. Time-domain, frequency-domain, and entropy-based multi-feature parameters are then extracted from these subtle feature components and concatenated into a multi-domain feature vector for feature extraction of the communication emitter’s symbol waveform. Finally, a long short-term memory (LSTM) network is employed to sequentially learn and classify the multi-domain feature vectors of each symbol of the emitter signal, thus achieving individual identification and classification of communication emitters. Experimental validation is conducted using the publicly available Oracle dataset. The results show that, when the signal-to-noise ratio (SNR) is 6 dB, the proposed algorithm achieves a recognition accuracy of 96.7%, representing a 22.1% improvement compared to the average reco gnition accuracy of individual domains.
Available online: June 13, 2025
,DOI: 10.15888/j.cnki.csa.009917
Abstract:
Urban remote sensing images pose challenges in boundary segmentation due to their high resolution, diverse backgrounds, and intricate textures. Mainstream semantic segmentation models encounter difficulties, including edge blurring, smooth corners, and the inability to capture long-range dependencies.To address these challenges, ARD-UNet++, an enhanced model based on UNet++, is introduced. A 7×7 depthwise separable convolution is employed to reduce the parameter count, facilitating denser feature extraction and comprehensive contextual information capture. The SimAM non-parametric attention mechanism is introduced to selectively focus on crucial features without introducing additional parameters, effectively suppressing irrelevant information. Residual connections are integrated to prevent local optima, with the Res-SimAMmodule replacing the standard convolution block in upsampling nodes. In comparison to UNet++, the proposedenhanced model demonstrates significant improvements in UAVid and Potsdam datasets, achieving a 6.77% and 1.79% increase in mIoU, 4.71% and 1.17% in F1, and 4.99% and 0.98% in OA, respectively. A comparative analysis against recent mainstream models underscores its superior performance, positioning ARD-UNet++ as a promising solution for precise urban remote sensing image segmentation.
Urban remote sensing images pose challenges in boundary segmentation due to their high resolution, diverse backgrounds, and intricate textures. Mainstream semantic segmentation models encounter difficulties, including edge blurring, smooth corners, and the inability to capture long-range dependencies.To address these challenges, ARD-UNet++, an enhanced model based on UNet++, is introduced. A 7×7 depthwise separable convolution is employed to reduce the parameter count, facilitating denser feature extraction and comprehensive contextual information capture. The SimAM non-parametric attention mechanism is introduced to selectively focus on crucial features without introducing additional parameters, effectively suppressing irrelevant information. Residual connections are integrated to prevent local optima, with the Res-SimAMmodule replacing the standard convolution block in upsampling nodes. In comparison to UNet++, the proposedenhanced model demonstrates significant improvements in UAVid and Potsdam datasets, achieving a 6.77% and 1.79% increase in mIoU, 4.71% and 1.17% in F1, and 4.99% and 0.98% in OA, respectively. A comparative analysis against recent mainstream models underscores its superior performance, positioning ARD-UNet++ as a promising solution for precise urban remote sensing image segmentation.
Available online: June 13, 2025
,DOI: 10.15888/j.cnki.csa.009920
Abstract:
Object-centric learning methods aim to parse and model scenes in a compositional way while extracting representations of objects within those scenes. Early object-centric approaches typically employ simple pixel-mixing decoders for scene modeling. However, these methods often perform poorly when handling complex synthetic datasets and real-world datasets. In contrast, recent object-centric learning methods have begun experimenting with more complex decoders, such as autoregressive Transformers and diffusion models, to extract object representations and model scenes more effectively. Despite the improved performance of these newer methods over earlier ones, their non-compositional modeling approaches contradict human intuition and fail to generate corresponding object images given object representations. To address this issue, the proposed object-centric diffusion (OCD) model employs an improved diffusion model as a decoder. OCD generates the appearance and masks of objects separately during the scene reconstruction process, achieving true compositional modeling while maintaining model performance. Extensive experiments demonstrate that OCD excels in image segmentation and generation tasks across various datasets, including two synthetic and two real-world datasets, proving its versatility and effectiveness.
Object-centric learning methods aim to parse and model scenes in a compositional way while extracting representations of objects within those scenes. Early object-centric approaches typically employ simple pixel-mixing decoders for scene modeling. However, these methods often perform poorly when handling complex synthetic datasets and real-world datasets. In contrast, recent object-centric learning methods have begun experimenting with more complex decoders, such as autoregressive Transformers and diffusion models, to extract object representations and model scenes more effectively. Despite the improved performance of these newer methods over earlier ones, their non-compositional modeling approaches contradict human intuition and fail to generate corresponding object images given object representations. To address this issue, the proposed object-centric diffusion (OCD) model employs an improved diffusion model as a decoder. OCD generates the appearance and masks of objects separately during the scene reconstruction process, achieving true compositional modeling while maintaining model performance. Extensive experiments demonstrate that OCD excels in image segmentation and generation tasks across various datasets, including two synthetic and two real-world datasets, proving its versatility and effectiveness.
Available online: May 12, 2025
,DOI: 10.15888/j.cnki.csa.009901
Abstract:
In response to the challenges of scene scale variability, intra-class diversity, and inter-class similarity in remote sensing image scene classification, a Vision Transformer (ViT) model that integrates multi-scale features with detail perception strategies is proposed for remote sensing image classification. The model effectively captures and fuses multi-scale features from remote sensing images by incorporating a dilated spatial pyramid pooling module, while enhancing the utilization of local feature information, thus improving feature discrimination capabilities. Furthermore, an innovative detail perception masking strategy enables the model to leverage unlabeled remote sensing image data effectively, facilitating the learning of more refined feature representations for more efficient and accurate scene classification. In the experimental section, the model is first pre-trained on a large-scale unlabeled remote sensing image dataset, followed by the fine-tuning of the pre-trained model on downstream scene classification tasks. Experimental results across multiple public remote sensing image datasets demonstrate that the proposed model can effectively extract image features during the self-supervised pre-training phase and achieve high accuracy in downstream scene classification tasks, showcasing robust performance and efficacy.
In response to the challenges of scene scale variability, intra-class diversity, and inter-class similarity in remote sensing image scene classification, a Vision Transformer (ViT) model that integrates multi-scale features with detail perception strategies is proposed for remote sensing image classification. The model effectively captures and fuses multi-scale features from remote sensing images by incorporating a dilated spatial pyramid pooling module, while enhancing the utilization of local feature information, thus improving feature discrimination capabilities. Furthermore, an innovative detail perception masking strategy enables the model to leverage unlabeled remote sensing image data effectively, facilitating the learning of more refined feature representations for more efficient and accurate scene classification. In the experimental section, the model is first pre-trained on a large-scale unlabeled remote sensing image dataset, followed by the fine-tuning of the pre-trained model on downstream scene classification tasks. Experimental results across multiple public remote sensing image datasets demonstrate that the proposed model can effectively extract image features during the self-supervised pre-training phase and achieve high accuracy in downstream scene classification tasks, showcasing robust performance and efficacy.
Available online: May 12, 2025
,DOI: 10.15888/j.cnki.csa.009895
Abstract:
Graphs play a crucial role in modeling relationships between entities across various applications. Workloads on graphs are typically categorized into transactional and analytical workloads. Many scenarios now require handling both types of workloads simultaneously. However, most existing graph storage systems are optimized for only one type of workload and cannot efficiently handle both simultaneously. In this study, a new graph storage system, HGraph, is proposed to address this issue. A data structure tailored for hybrid workloads is designed through careful analysis of the access patterns of both workload types. In addition, HGraph introduces a multi-version concurrency control (MVCC) implementation based on undo logs, which is memory-efficient and improves traversal performance. HGraph also adopts copy-on-write and optimistic concurrency control strategies to optimize transaction processing, further enhancing system concurrency. Extensive experiments on both real-world and synthetic datasets demonstrate that HGraph outperforms other graph storage systems.
Graphs play a crucial role in modeling relationships between entities across various applications. Workloads on graphs are typically categorized into transactional and analytical workloads. Many scenarios now require handling both types of workloads simultaneously. However, most existing graph storage systems are optimized for only one type of workload and cannot efficiently handle both simultaneously. In this study, a new graph storage system, HGraph, is proposed to address this issue. A data structure tailored for hybrid workloads is designed through careful analysis of the access patterns of both workload types. In addition, HGraph introduces a multi-version concurrency control (MVCC) implementation based on undo logs, which is memory-efficient and improves traversal performance. HGraph also adopts copy-on-write and optimistic concurrency control strategies to optimize transaction processing, further enhancing system concurrency. Extensive experiments on both real-world and synthetic datasets demonstrate that HGraph outperforms other graph storage systems.
Available online: May 12, 2025
,DOI: 10.15888/j.cnki.csa.009897
Abstract:
Due to the scarcity of defective fabric samples in real-world applications and the subtle nature of anomaly regions, existing methods may fail to detect defect areas accurately, leading to localization errors. To leverage the Transformer attention mechanism’s ability to capture fine-grained features, this study proposes a dynamic attention-sharing mechanism tailored for fabric defect detection tasks, enhancing the model’s ability to capture subtle textures. Specifically, a fabric defect detection framework based on a mix multi-head attention (MMHA) reconstruction network is proposed. First, a pre-trained Transformer encoder is used to extract features from fabric images. Next, a bottleneck layer with MeanDropout is employed to reduce the model’s reliance on redundant features. Furthermore, a decoder combining MMHA with LLamaMLP is introduced to facilitate the dynamic selection of relevant fabric features, thus improving attention to critical fine-grained textures. Finally, the decoder loosely reconstructs combinations of multi-level features to achieve defect detection and localization. Extensive experiments conducted on publicly available datasets demonstrate that the proposed method achieves state-of-the-art performance in fabric defect detection. Experimental results show that the proposed method improves image-level indicators by 2.1% and 0.6% on two datasets, respectively, and achieves leading performance with pixel-level indicators of 96.0% and 60.5%.
Due to the scarcity of defective fabric samples in real-world applications and the subtle nature of anomaly regions, existing methods may fail to detect defect areas accurately, leading to localization errors. To leverage the Transformer attention mechanism’s ability to capture fine-grained features, this study proposes a dynamic attention-sharing mechanism tailored for fabric defect detection tasks, enhancing the model’s ability to capture subtle textures. Specifically, a fabric defect detection framework based on a mix multi-head attention (MMHA) reconstruction network is proposed. First, a pre-trained Transformer encoder is used to extract features from fabric images. Next, a bottleneck layer with MeanDropout is employed to reduce the model’s reliance on redundant features. Furthermore, a decoder combining MMHA with LLamaMLP is introduced to facilitate the dynamic selection of relevant fabric features, thus improving attention to critical fine-grained textures. Finally, the decoder loosely reconstructs combinations of multi-level features to achieve defect detection and localization. Extensive experiments conducted on publicly available datasets demonstrate that the proposed method achieves state-of-the-art performance in fabric defect detection. Experimental results show that the proposed method improves image-level indicators by 2.1% and 0.6% on two datasets, respectively, and achieves leading performance with pixel-level indicators of 96.0% and 60.5%.
Available online: April 25, 2025
,DOI: 10.15888/j.cnki.csa.009900
Abstract:
In the field of deep reinforcement learning, particularly for high-dimensional continuous tasks, efficiently utilizing limited training data, preventing overfitting, and enhancing the model’s generalization ability are crucial research challenges. Traditional reinforcement learning algorithms typically rely on a single experience replay buffer, which, when applied to high-dimensional continuous state and action spaces, often faces low exploration efficiency and insufficient sample utilization. A reinforcement learning experience replay mechanism based on sample distinctiveness called distinctive experience replay (DER) is proposed. This mechanism selects samples with notable distinctiveness for experience replay. The core concept of DER is to identify and select significantly distinctive samples during training and store them in a dedicated experience pool. This mechanism not only effectively utilizes diverse samples to prevent neural network overfitting but also enhances the agent’s learning efficiency and decision-making quality in complex environments. Experimental results show that DER significantly improves the agent’s learning efficiency and final performance in classic reinforcement learning environments.
In the field of deep reinforcement learning, particularly for high-dimensional continuous tasks, efficiently utilizing limited training data, preventing overfitting, and enhancing the model’s generalization ability are crucial research challenges. Traditional reinforcement learning algorithms typically rely on a single experience replay buffer, which, when applied to high-dimensional continuous state and action spaces, often faces low exploration efficiency and insufficient sample utilization. A reinforcement learning experience replay mechanism based on sample distinctiveness called distinctive experience replay (DER) is proposed. This mechanism selects samples with notable distinctiveness for experience replay. The core concept of DER is to identify and select significantly distinctive samples during training and store them in a dedicated experience pool. This mechanism not only effectively utilizes diverse samples to prevent neural network overfitting but also enhances the agent’s learning efficiency and decision-making quality in complex environments. Experimental results show that DER significantly improves the agent’s learning efficiency and final performance in classic reinforcement learning environments.
-
全文下载排行(总排行，年度排行，各期排行)
摘要点击排行(总排行，年度排行，各期排行)
Article Search
Search by issue
Select AllDeselectExport
Display Method:
2016,25(8):1-7
,DOI: 10.15888/j.cnki.csa.005283
Abstract:
Since 2006, Deep Neural Network has achieved huge access in the area of Big Data Processing and Artificial Intelligence, such as image/video discriminations and autopilot. And unsupervised learning methods as the methods getting success in the depth neural network pre training play an important role in deep learning. So, this paper attempts to make a brief introduction and analysis of unsupervised learning methods in deep learning, mainly includs two types, Auto-Encoders based on determination theory and Contrastive Divergence for Restrict Boltzmann Machine based on probability theory. Secondly, the applications of the two methods in Deep Learning are introduced. At last a brief summary and prospect of the challenges faced by unsupervised learning methods in Deep Neural Networks are made.
Since 2006, Deep Neural Network has achieved huge access in the area of Big Data Processing and Artificial Intelligence, such as image/video discriminations and autopilot. And unsupervised learning methods as the methods getting success in the depth neural network pre training play an important role in deep learning. So, this paper attempts to make a brief introduction and analysis of unsupervised learning methods in deep learning, mainly includs two types, Auto-Encoders based on determination theory and Contrastive Divergence for Restrict Boltzmann Machine based on probability theory. Secondly, the applications of the two methods in Deep Learning are introduced. At last a brief summary and prospect of the challenges faced by unsupervised learning methods in Deep Neural Networks are made.
2008,17(5):122-126
Abstract:
随着Internet的迅速发展，网络资源越来越丰富，人们如何从网络上抽取信息也变得至关重要，尤其是占网络资源80%的Deep Web信息检索更是人们应该倍加关注的难点问题。为了更好的研究Deep Web爬虫技术，本文对有关Deep Web爬虫的内容进行了全面、详细地介绍。首先对Deep Web爬虫的定义及研究目标进行了阐述，接着介绍了近年来国内外关于Deep Web爬虫的研究进展，并对其加以分析。在此基础上展望了Deep Web爬虫的研究趋势，为下一步的研究奠定了基础。
随着Internet的迅速发展，网络资源越来越丰富，人们如何从网络上抽取信息也变得至关重要，尤其是占网络资源80%的Deep Web信息检索更是人们应该倍加关注的难点问题。为了更好的研究Deep Web爬虫技术，本文对有关Deep Web爬虫的内容进行了全面、详细地介绍。首先对Deep Web爬虫的定义及研究目标进行了阐述，接着介绍了近年来国内外关于Deep Web爬虫的研究进展，并对其加以分析。在此基础上展望了Deep Web爬虫的研究趋势，为下一步的研究奠定了基础。
2011,20(11):80-85
Abstract:
Based on study of current video transcoding solutions, we proposed a distributed transcoding system. Video resources are stored in HDFS(Hadoop Distributed File System) and transcoded by MapReduce program using FFMPEG. In this paper, video segmentation strategy on distributed storage and how they affect accessing time are discussed. We also defined metadata of video formats and transcoding parameters. The distributed transcoding framework is proposed on basis of MapReduce programming model. Segmented source videos are transcoding in map tasks and merged into target video in reduce task. Experimental results show that transcoding time is dependent on segmentation size and trascoding cluster size. Compared with single PC, the proposed distributed video transcoding system implemented on 8 PCs can decrease about 80% of the transcoding time.
Based on study of current video transcoding solutions, we proposed a distributed transcoding system. Video resources are stored in HDFS(Hadoop Distributed File System) and transcoded by MapReduce program using FFMPEG. In this paper, video segmentation strategy on distributed storage and how they affect accessing time are discussed. We also defined metadata of video formats and transcoding parameters. The distributed transcoding framework is proposed on basis of MapReduce programming model. Segmented source videos are transcoding in map tasks and merged into target video in reduce task. Experimental results show that transcoding time is dependent on segmentation size and trascoding cluster size. Compared with single PC, the proposed distributed video transcoding system implemented on 8 PCs can decrease about 80% of the transcoding time.
2022,31(5):1-20
,DOI: 10.15888/j.cnki.csa.008463
Abstract:
Although the deep learning method has made a huge breakthrough in machine learning, it requires a large amount of manual work for data annotation. Limited by labor costs, however, many applications are expected to reason and judge the instance labels that have never been encountered before. For this reason, zero-shot learning (ZSL) came into being. As a natural data structure that represents the connection between things, the graph is currently drawing more and more attention in ZSL. Therefore, this study reviews the methods of graph-based ZSL systematically. Firstly, the definitions of ZSL and graph learning are outlined, and the ideas of existing solutions for ZSL are summarized. Secondly, the current ZSL methods are classified according to different utilization ways of graphs. Thirdly, the evaluation criteria and datasets concerning graph-based ZSL are discussed. Finally, this study also specifies the problems to be solved in further research on graph-based ZSL and predicts the possible directions of its future development.
Although the deep learning method has made a huge breakthrough in machine learning, it requires a large amount of manual work for data annotation. Limited by labor costs, however, many applications are expected to reason and judge the instance labels that have never been encountered before. For this reason, zero-shot learning (ZSL) came into being. As a natural data structure that represents the connection between things, the graph is currently drawing more and more attention in ZSL. Therefore, this study reviews the methods of graph-based ZSL systematically. Firstly, the definitions of ZSL and graph learning are outlined, and the ideas of existing solutions for ZSL are summarized. Secondly, the current ZSL methods are classified according to different utilization ways of graphs. Thirdly, the evaluation criteria and datasets concerning graph-based ZSL are discussed. Finally, this study also specifies the problems to be solved in further research on graph-based ZSL and predicts the possible directions of its future development.
2012,21(3):260-264
Abstract:
The essential problem of open platform is the validation and authorization of users. Nowadays, OAuth is the international authorization method. Its characteristic is that users could apply to visit their protected resources without the need to enter their names and passwords in the third application. The latest version of OAuth is OAuth2.0 and its operation of validation and authorization are simpler and safer. This paper investigates the principle of OAuth2.0, analyzes the procedure of Refresh Token and offers a design proposal of OAuth2.0 server and specific application examples.
The essential problem of open platform is the validation and authorization of users. Nowadays, OAuth is the international authorization method. Its characteristic is that users could apply to visit their protected resources without the need to enter their names and passwords in the third application. The latest version of OAuth is OAuth2.0 and its operation of validation and authorization are simpler and safer. This paper investigates the principle of OAuth2.0, analyzes the procedure of Refresh Token and offers a design proposal of OAuth2.0 server and specific application examples.
2007,16(9):22-25
Abstract:
本文结合物流遗留系统的实际安全状态,分析了面向对象的编程思想在横切关注点和核心关注点处理上的不足,指出面向方面的编程思想解决方案对系统进行分离关注点处理的优势,并对面向方面的编程的一种具体实现AspectJ进行分析,提出了一种依据AspectJ对遗留物流系统进行IC卡安全进化的方法.
本文结合物流遗留系统的实际安全状态,分析了面向对象的编程思想在横切关注点和核心关注点处理上的不足,指出面向方面的编程思想解决方案对系统进行分离关注点处理的优势,并对面向方面的编程的一种具体实现AspectJ进行分析,提出了一种依据AspectJ对遗留物流系统进行IC卡安全进化的方法.
2002,11(10):70-73
Abstract:
2011,20(7):184-187,120
Abstract:
According to the actual needs of intelligent household, environmental monitoring etc, this paper designed a wireless sensor node of long-distance communication system. This system used the second SoC CC2530 set in RF and controller chips as the core module and externally connected with CC2591 RF front-end power amplifier module. Based on ZigBee2006 in software agreement stack, it realized each application layer function based on ZStack. It also introduced wireless data acquisition networks based on the ZigBee agreement construction, and has given the hardware design schematic diagram and the software flow chart of sensor node, synchronizer node. The experiment proved that the node is good in performance and the communication is reliable. The communication distance has increased obviously compared with the first generation TI product.
According to the actual needs of intelligent household, environmental monitoring etc, this paper designed a wireless sensor node of long-distance communication system. This system used the second SoC CC2530 set in RF and controller chips as the core module and externally connected with CC2591 RF front-end power amplifier module. Based on ZigBee2006 in software agreement stack, it realized each application layer function based on ZStack. It also introduced wireless data acquisition networks based on the ZigBee agreement construction, and has given the hardware design schematic diagram and the software flow chart of sensor node, synchronizer node. The experiment proved that the node is good in performance and the communication is reliable. The communication distance has increased obviously compared with the first generation TI product.
2019,28(6):1-12
,DOI: 10.15888/j.cnki.csa.006915
Abstract:
A knowledge graph is a knowledge base that represents objective concepts/entities and their relationships in the form of graph, which is one of the fundamental technologies for intelligent services such as semantic retrieval, intelligent answering, decision support, etc. Currently, the connotation of knowledge graph is not clear enough and the usage/reuse rate of existing knowledge graphs is relatively low due to lack of documentation. This paper clarifies the concept of knowledge graph through differentiating it from related concepts such as ontology in that the ontology is the schema layer and the logical basis of a knowledge graph while the knowledge graph is the instantiation of an ontology. Research results of ontologies can be used as the foundation of knowledge graph research to promote its developments and applications. Existing generic/domain knowledge graphs are briefly documented and analyzed in terms of building, storage, and retrieval methods. Moreover, future research directions are pointed out.
A knowledge graph is a knowledge base that represents objective concepts/entities and their relationships in the form of graph, which is one of the fundamental technologies for intelligent services such as semantic retrieval, intelligent answering, decision support, etc. Currently, the connotation of knowledge graph is not clear enough and the usage/reuse rate of existing knowledge graphs is relatively low due to lack of documentation. This paper clarifies the concept of knowledge graph through differentiating it from related concepts such as ontology in that the ontology is the schema layer and the logical basis of a knowledge graph while the knowledge graph is the instantiation of an ontology. Research results of ontologies can be used as the foundation of knowledge graph research to promote its developments and applications. Existing generic/domain knowledge graphs are briefly documented and analyzed in terms of building, storage, and retrieval methods. Moreover, future research directions are pointed out.
2008,17(1):113-116
Abstract:
排序是计算机程序设计中一种重要操作，本文论述了C语言中快速排序算法的改进，即快速排序与直接插入排序算法相结合的实现过程。在C语言程序设计中，实现大量的内部排序应用时，所寻求的目的就是找到一个简单、有效、快捷的算法。本文着重阐述快速排序的改进与提高过程，从基本的性能特征到基本的算法改进，通过不断的分析，实验，最后得出最佳的改进算法。
排序是计算机程序设计中一种重要操作，本文论述了C语言中快速排序算法的改进，即快速排序与直接插入排序算法相结合的实现过程。在C语言程序设计中，实现大量的内部排序应用时，所寻求的目的就是找到一个简单、有效、快捷的算法。本文着重阐述快速排序的改进与提高过程，从基本的性能特征到基本的算法改进，通过不断的分析，实验，最后得出最佳的改进算法。
2008,17(8):87-89
Abstract:
随着面向对象软件开发技术的广泛应用和软件测试自动化的要求，基于模型的软件测试逐渐得到了软件开发人员和软件测试人员的认可和接受。基于模型的软件测试是软件编码阶段的主要测试方法之一，具有测试效率高、排除逻辑复杂故障测试效果好等特点。但是误报、漏报和故障机理有待进一步研究。对主要的测试模型进行了分析和分类，同时，对故障密度等参数进行了初步的分析；最后，提出了一种基于模型的软件测试流程。
随着面向对象软件开发技术的广泛应用和软件测试自动化的要求，基于模型的软件测试逐渐得到了软件开发人员和软件测试人员的认可和接受。基于模型的软件测试是软件编码阶段的主要测试方法之一，具有测试效率高、排除逻辑复杂故障测试效果好等特点。但是误报、漏报和故障机理有待进一步研究。对主要的测试模型进行了分析和分类，同时，对故障密度等参数进行了初步的分析；最后，提出了一种基于模型的软件测试流程。
2008,17(8):2-5
Abstract:
本文介绍了一个企业信息门户中单点登录系统的设计与实现。系统实现了一个基于Java EE架构的结合凭证加密和Web Services的单点登录系统，对门户用户进行统一认证和访问控制。论文详细阐述了该系统的总体结构、设计思想、工作原理和具体实现方案，目前系统已在部分省市的广电行业信息门户平台中得到了良好的应用。
本文介绍了一个企业信息门户中单点登录系统的设计与实现。系统实现了一个基于Java EE架构的结合凭证加密和Web Services的单点登录系统，对门户用户进行统一认证和访问控制。论文详细阐述了该系统的总体结构、设计思想、工作原理和具体实现方案，目前系统已在部分省市的广电行业信息门户平台中得到了良好的应用。
2004,13(8):58-59
Abstract:
本文介绍了Visual C++6.0在对话框的多个文本框之间,通过回车键转移焦点的几种方法,并提出了一个改进方法.
本文介绍了Visual C++6.0在对话框的多个文本框之间,通过回车键转移焦点的几种方法,并提出了一个改进方法.
2009,18(5):182-185
Abstract:
DICOM 是医学图像存储和传输的国际标准，DCMTK 是免费开源的针对DICOM 标准的开发包。解读DICOM 文件格式并解决DICOM 医学图像显示问题是医学图像处理的基础，对医学影像技术的研究具有重要意义。解读了DICOM 文件格式并介绍了调窗处理的原理，利用VC++和DCMTK 实现医学图像显示和调窗功能。
DICOM 是医学图像存储和传输的国际标准，DCMTK 是免费开源的针对DICOM 标准的开发包。解读DICOM 文件格式并解决DICOM 医学图像显示问题是医学图像处理的基础，对医学影像技术的研究具有重要意义。解读了DICOM 文件格式并介绍了调窗处理的原理，利用VC++和DCMTK 实现医学图像显示和调窗功能。
-
全文下载排行(总排行，年度排行，各期排行)
摘要点击排行(总排行，年度排行，各期排行)
Article Search
Search by issue
Select AllDeselectExport
Display Method:
2007,16(10):48-51
Abstract:
论文对HDF数据格式和函数库进行研究,重点以栅格图像为例,详细论述如何利用VC++.net和VC#.net对光栅数据进行读取与处理,然后根据所得到的象素矩阵用描点法显示图像.论文是以国家气象中心开发Micaps3.0(气象信息综合分析处理系统)的课题研究为背景的.
论文对HDF数据格式和函数库进行研究,重点以栅格图像为例,详细论述如何利用VC++.net和VC#.net对光栅数据进行读取与处理,然后根据所得到的象素矩阵用描点法显示图像.论文是以国家气象中心开发Micaps3.0(气象信息综合分析处理系统)的课题研究为背景的.
2002,11(12):67-68
Abstract:
本文介绍非实时操作系统Windows 2000下,利用VisualC++6.0开发实时数据采集的方法.所用到的数据采集卡是研华的PCL-818L.借助数据采集卡PCL-818L的DLLs中的API函数,提出三种实现高速实时数据采集的方法及优缺点.
本文介绍非实时操作系统Windows 2000下,利用VisualC++6.0开发实时数据采集的方法.所用到的数据采集卡是研华的PCL-818L.借助数据采集卡PCL-818L的DLLs中的API函数,提出三种实现高速实时数据采集的方法及优缺点.
2008,17(1):113-116
Abstract:
排序是计算机程序设计中一种重要操作，本文论述了C语言中快速排序算法的改进，即快速排序与直接插入排序算法相结合的实现过程。在C语言程序设计中，实现大量的内部排序应用时，所寻求的目的就是找到一个简单、有效、快捷的算法。本文着重阐述快速排序的改进与提高过程，从基本的性能特征到基本的算法改进，通过不断的分析，实验，最后得出最佳的改进算法。
排序是计算机程序设计中一种重要操作，本文论述了C语言中快速排序算法的改进，即快速排序与直接插入排序算法相结合的实现过程。在C语言程序设计中，实现大量的内部排序应用时，所寻求的目的就是找到一个简单、有效、快捷的算法。本文着重阐述快速排序的改进与提高过程，从基本的性能特征到基本的算法改进，通过不断的分析，实验，最后得出最佳的改进算法。
2008,17(5):122-126
Abstract:
随着Internet的迅速发展，网络资源越来越丰富，人们如何从网络上抽取信息也变得至关重要，尤其是占网络资源80%的Deep Web信息检索更是人们应该倍加关注的难点问题。为了更好的研究Deep Web爬虫技术，本文对有关Deep Web爬虫的内容进行了全面、详细地介绍。首先对Deep Web爬虫的定义及研究目标进行了阐述，接着介绍了近年来国内外关于Deep Web爬虫的研究进展，并对其加以分析。在此基础上展望了Deep Web爬虫的研究趋势，为下一步的研究奠定了基础。
随着Internet的迅速发展，网络资源越来越丰富，人们如何从网络上抽取信息也变得至关重要，尤其是占网络资源80%的Deep Web信息检索更是人们应该倍加关注的难点问题。为了更好的研究Deep Web爬虫技术，本文对有关Deep Web爬虫的内容进行了全面、详细地介绍。首先对Deep Web爬虫的定义及研究目标进行了阐述，接着介绍了近年来国内外关于Deep Web爬虫的研究进展，并对其加以分析。在此基础上展望了Deep Web爬虫的研究趋势，为下一步的研究奠定了基础。
2025,34(7):1-13, DOI: 10.15888/j.cnki.csa.009942
Abstract:
2025,34(6):1-11, DOI: 10.15888/j.cnki.csa.009903
Abstract:
2025,34(5):1-13, DOI: 10.15888/j.cnki.csa.009879
Abstract:
2025,34(4):18-33, DOI: 10.15888/j.cnki.csa.009865
Abstract:
2025,34(4):1-17, DOI: 10.15888/j.cnki.csa.009839
Abstract:
2025,34(1):1-10, DOI: 10.15888/j.cnki.csa.009782
Abstract:
2025,34(1):26-36, DOI: 10.15888/j.cnki.csa.009733
Abstract:
2025,34(1):11-25, DOI: 10.15888/j.cnki.csa.009709
Abstract:
2024,33(12):1-15, DOI: 10.15888/j.cnki.csa.009708
Abstract:
2024,33(11):38-47, DOI: 10.15888/j.cnki.csa.009702
Abstract:
2024,33(10):26-36, DOI: 10.15888/j.cnki.csa.009646
Abstract:
2024,33(6):16-27, DOI: 10.15888/j.cnki.csa.009559
Abstract:
2024,33(6):1-15, DOI: 10.15888/j.cnki.csa.009531
Abstract:
2024,33(4):50-59, DOI: 10.15888/j.cnki.csa.009486
Abstract:
Application Review of Convolutional Neural Networks in Pathological Images Diagnosis of Liver Cancer
2024,33(4):26-38, DOI: 10.15888/j.cnki.csa.009466
Abstract:
2024,33(3):1-11, DOI: 10.15888/j.cnki.csa.009433
Abstract:
2024,33(3):12-23, DOI: 10.15888/j.cnki.csa.009421
Abstract:
2024,33(1):37-48, DOI: 10.15888/j.cnki.csa.009377
Abstract:
2024,33(1):22-36, DOI: 10.15888/j.cnki.csa.009369
Abstract:
2023,32(12):32-42, DOI: 10.15888/j.cnki.csa.009356
Abstract:
2023,32(11):36-47, DOI: 10.15888/j.cnki.csa.009269
Abstract:
2023,32(9):1-18, DOI: 10.15888/j.cnki.csa.009226
Abstract:
2023,32(8):1-18, DOI: 10.15888/j.cnki.csa.009208
Abstract:
2023,32(4):1-15, DOI: 10.15888/j.cnki.csa.009038
Abstract:
2023,32(4):16-41, DOI: 10.15888/j.cnki.csa.009037
Abstract:
2023,32(3):25-47, DOI: 10.15888/j.cnki.csa.008980
Abstract:
2023,32(3):1-14, DOI: 10.15888/j.cnki.csa.008976
Abstract:
2023,32(2):1-12, DOI: 10.15888/j.cnki.csa.008743
Abstract:
2023,32(1):12-28, DOI: 10.15888/j.cnki.csa.008914
Abstract:
2023,32(1):29-40, DOI: 10.15888/j.cnki.csa.008909
Abstract:
2023,32(1):1-11, DOI: 10.15888/j.cnki.csa.008848
Abstract:
2023,32(1):41-49, DOI: 10.15888/j.cnki.csa.008805
Abstract:
2022,31(12):20-28, DOI: 10.15888/j.cnki.csa.008838
Abstract:
2022,31(12):1-9, DOI: 10.15888/j.cnki.csa.008825
Abstract:
2022,31(11):49-59, DOI: 10.15888/j.cnki.csa.008847
Abstract:
2022,31(11):31-48, DOI: 10.15888/j.cnki.csa.008811
Abstract:
2022,31(10):44-50, DOI: 10.15888/j.cnki.csa.008776
Abstract:
2022,31(10):25-35, DOI: 10.15888/j.cnki.csa.008734
Abstract:
2022,31(10):36-43, DOI: 10.15888/j.cnki.csa.008711
Abstract:
2022,31(9):57-69, DOI: 10.15888/j.cnki.csa.008661
Abstract:
2022,31(8):1-16, DOI: 10.15888/j.cnki.csa.008713
Abstract:
2022,31(7):12-22, DOI: 10.15888/j.cnki.csa.008641
Abstract:
2022,31(7):35-45, DOI: 10.15888/j.cnki.csa.008595
Abstract:
2022,31(7):46-54, DOI: 10.15888/j.cnki.csa.008590
Abstract:
2022,31(7):23-34, DOI: 10.15888/j.cnki.csa.008566
Abstract:
2022,31(7):1-11, DOI: 10.15888/j.cnki.csa.008545
Abstract:
2022,31(5):40-51, DOI: 10.15888/j.cnki.csa.008476
Abstract:
2022,31(5):1-20, DOI: 10.15888/j.cnki.csa.008463
Abstract:
2022,31(5):21-29, DOI: 10.15888/j.cnki.csa.008426
Abstract:
2022,31(4):14-32, DOI: 10.15888/j.cnki.csa.008419
Abstract:
2022,31(4):1-13, DOI: 10.15888/j.cnki.csa.008418
Abstract:
2022,31(4):47-58, DOI: 10.15888/j.cnki.csa.008403
Abstract:
2022,31(3):1-8, DOI: 10.15888/j.cnki.csa.008415
Abstract:
2022,31(3):48-55, DOI: 10.15888/j.cnki.csa.008377
Abstract:
2022,31(2):13-21, DOI: 10.15888/j.cnki.csa.008319
Abstract:
2022,31(2):1-12, DOI: 10.15888/j.cnki.csa.008303
Abstract:
2022,31(1):1-10, DOI: 10.15888/j.cnki.csa.008323
Abstract:
2021,30(12):28-36, DOI: 10.15888/j.cnki.csa.008289
Abstract:
2021,30(12):10-17, DOI: 10.15888/j.cnki.csa.008252
Abstract:
2021,30(12):1-9, DOI: 10.15888/j.cnki.csa.008245
Abstract:
2021,30(10):21-30, DOI: 10.15888/j.cnki.csa.008137
Abstract:
2021,30(9):12-23, DOI: 10.15888/j.cnki.csa.008086
Abstract:
2021,30(9):24-40, DOI: 10.15888/j.cnki.csa.008079
Abstract:
2021,30(9):41-52, DOI: 10.15888/j.cnki.csa.008078
Abstract:
2021,30(8):14-21, DOI: 10.15888/j.cnki.csa.008003
Abstract:
2021,30(7):13-21, DOI: 10.15888/j.cnki.csa.007992
Abstract:
2021,30(7):1-12, DOI: 10.15888/j.cnki.csa.007973
Abstract:
2021,30(6):18-27, DOI: 10.15888/j.cnki.csa.007956
Abstract:
2021,30(5):21-30, DOI: 10.15888/j.cnki.csa.007916
Abstract:
2021,30(3):1-13, DOI: 10.15888/j.cnki.csa.007839
Abstract:
2021,30(1):10-18, DOI: 10.15888/j.cnki.csa.007776
Abstract:
2020,29(12):13-25, DOI: 10.15888/j.cnki.csa.007701
Abstract:
2020,29(11):29-39, DOI: 10.15888/j.cnki.csa.007705
Abstract:
2020,29(9):16-25, DOI: 10.15888/j.cnki.csa.007632
Abstract:
2020,29(9):26-31, DOI: 10.15888/j.cnki.csa.007617
Abstract:
2020,29(9):1-15, DOI: 10.15888/j.cnki.csa.007489
Abstract:
2020,29(8):1-8, DOI: 10.15888/j.cnki.csa.007478
Abstract:
2020,29(6):56-63, DOI: 10.15888/j.cnki.csa.007460
Abstract:
2020,29(6):1-12, DOI: 10.15888/j.cnki.csa.007431
Abstract:
2020,29(6):13-21, DOI: 10.15888/j.cnki.csa.007430
Abstract:
2020,29(3):29-38, DOI: 10.15888/j.cnki.csa.007315
Abstract:
2020,29(3):1-10, DOI: 10.15888/j.cnki.csa.007301
Abstract:
2020,29(3):11-19, DOI: 10.15888/j.cnki.csa.007261
Abstract:
2020,29(2):1-8, DOI: 10.15888/j.cnki.csa.007275
Abstract:
2020,29(2):9-27, DOI: 10.15888/j.cnki.csa.007262
Abstract:
2020,29(1):14-21, DOI: 10.15888/j.cnki.csa.007243
Abstract:
2020,29(1):22-28, DOI: 10.15888/j.cnki.csa.007237
Abstract:
2019,28(12):1-8, DOI: 10.15888/j.cnki.csa.007195
Abstract:
2019,28(11):1-9, DOI: 10.15888/j.cnki.csa.007156
Abstract:
2019,28(9):1-8, DOI: 10.15888/j.cnki.csa.007037
Abstract:
2019,28(7):1-8, DOI: 10.15888/j.cnki.csa.006988
Abstract:
2019,28(6):1-12, DOI: 10.15888/j.cnki.csa.006915
Abstract:
2019,28(4):205-211, DOI: 10.15888/j.cnki.csa.006739
Abstract:
2019,28(3):1-9, DOI: 10.15888/j.cnki.csa.006812
Abstract:
2019,28(3):10-17, DOI: 10.15888/j.cnki.csa.006806
Abstract:
2019,28(1):1-9, DOI: 10.15888/j.cnki.csa.006720
Abstract:
2019,28(1):10-16, DOI: 10.15888/j.cnki.csa.006707
Abstract:
2018,27(10):39-45, DOI: 10.15888/j.cnki.csa.006591
Abstract:
2018,27(10):33-38, DOI: 10.15888/j.cnki.csa.006560
Abstract:
2018,27(9):47-51, DOI: 10.15888/j.cnki.csa.006542
Abstract:
2018,27(8):10-18, DOI: 10.15888/j.cnki.csa.006456
Abstract:
2018,27(7):1-10, DOI: 10.15888/j.cnki.csa.006451
Abstract:
2018,27(5):1-9, DOI: 10.15888/j.cnki.csa.006336
Abstract:
2018,27(4):1-9, DOI: 10.15888/j.cnki.csa.006273
Abstract:
2018,27(1):1-9, DOI: 10.15888/j.cnki.csa.006169
Abstract:
2017,26(12):32-36, DOI: 10.15888/j.cnki.csa.006129
Abstract:
2017,26(12):25-31, DOI: 10.15888/j.cnki.csa.006096
Abstract:
2017,26(6):36-39, DOI: 10.15888/j.cnki.csa.005807
Abstract:
2017,26(4):1-8, DOI: 10.15888/j.cnki.csa.005773
Abstract:
2017,26(2):30-36, DOI: 10.15888/j.cnki.csa.005626
Abstract:
2016,25(9):17-26, DOI: 10.15888/j.cnki.csa.005354
Abstract:
2016,25(9):10-16, DOI: 10.15888/j.cnki.csa.005330
Abstract:
2016,25(8):1-7, DOI: 10.15888/j.cnki.csa.005283
Abstract:
2016,25(6):9-16, DOI: 10.15888/j.cnki.csa.005173
Abstract:
2016,25(6):1-8, DOI: 10.15888/j.cnki.csa.005206
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
Abstract:
下载归智APP ，关注本刊
External Links
You are the firstVisitors
Copyright: Institute of Software, Chinese Academy of Sciences Beijing ICP No. 05046678-3
Address：4# South Fourth Street, Zhongguancun,Haidian, Beijing,Postal Code：100190
Phone：010-62661041 Fax： Email：csa (a) iscas.ac.cn
Technical Support：Beijing Qinyun Technology Development Co., Ltd.
Copyright: Institute of Software, Chinese Academy of Sciences Beijing ICP No. 05046678-3
Address：4# South Fourth Street, Zhongguancun,Haidian, Beijing,Postal Code：100190
Phone：010-62661041 Fax： Email：csa (a) iscas.ac.cn
Technical Support：Beijing Qinyun Technology Development Co., Ltd.
