2023, 32(5):1-10. DOI: 10.15888/j.cnki.csa.009086
Abstract:Lack of concentration is an attention disorder that is common among teenagers, and it directly affects people’s learning and work efficiency. Most of the traditional attention detection methods rely on the observation of expressions, postures, and other behaviors and fail to objectively and accurately reflect attention states. Amid the rapid development of physiological detection technology, attention detection based on electroencephalography (EEG) signals has received considerable attention recently. However, related studies still have the problem of low detection accuracy. In this study, the EEG signals of 155 college students in the three states of being focused, distracted, and relaxed are collected, and the three attention states are identified by various machine learning methods on the basis of the wavelet features, differential entropy features and power spectrum features of the signals. The results show that these features of EEG signals can effectively distinguish the attention states of the subjects. The average accuracy of the detection method based on symmetrical dual-channel features is (80.84±3)%, and the detection precision of this method is significantly higher than that of the method based on single-channel features.
2023, 32(5):11-19. DOI: 10.15888/j.cnki.csa.009063
Abstract:Relief shading is an important part of large-scale battlefield simulation. Aiming at the problem that texture features of the existing relief shading technologies are not obvious in terms of details, this study proposes a large-scale battlefield relief shading enhancement method that combines elevation curvature and ambient occlusion. In the first step, by analyzing the curvature attribute of digital elevation data, a terrain curvature map is generated and then superimposed with satellite images to highlight geomorphic feature lines. In the second step, an ambient occlusion calculation method based on depthwise separable convolution is proposed, which can enhance the visual performance of battlefield terrain in gullies. Finally, the curvature map, ambient occlusion, and satellite images are fused to generate a real-time relief shading effect. Experiments show that the proposed method can present better visual effects on low-level global satellite images so that the observer can further analyze the texture features in terms of terrain details while grasping the overall trend of the three-dimensional terrain.
2023, 32(5):20-27. DOI: 10.15888/j.cnki.csa.009062
Abstract:A lightweight network based on depthwise separable convolution and the attention mechanism is proposed for fast detection of surface defects on semiconductor wafers, and experiments are conducted on the WM-811K dataset. As the proportions of defects of nine different categories in this dataset are imbalanced, a data enhancement method is used to expand the data for defect categories with few data. The depthwise separable convolution in this model can reduce the number of parameters and improve the inference speed of the model. The attention mechanism can make the model pay more attention to the defective regions in the wafer image so that the model can achieve better classification results. The experiments show that the average accuracy of the proposed method on the WM-811K dataset is as high as 96.5%, which is improved to varying degrees compared with that of ANN, VGG16, and MobileNetv2. In addition, the number of parameters and the amount of operation are only 73.5% and 28.6% of those of the classical lightweight network MobileNetv2, respectively.
2023, 32(5):28-35. DOI: 10.15888/j.cnki.csa.009120
Abstract:In recent years, researchers have found that the hyperspectral image classification method based on dual branch structure can more effectively extract the spectral and spatial features of the image for classification. However, in the dual branch structure, each branch only focuses on refining and extracting spectral or spatial features, with the study on cross-dimensional spectral-spatial feature interaction ignored, and the partial interaction extracted by the two branches respectively is not obvious, which affects the performance of classification. To solve this problem, this study proposes a hyperspectral image classification method based on global attention information interaction. First, the dense connection network is used to divide the image into two branches to refine the spectral and spatial features, respectively, and then the channel global attention features and spatial global attention features are obtained by combining the global attention mechanism (GAM). Finally, an information interaction module is used to realize the interaction of spectral and spatial information, which makes full use of spectral and spatial information to achieve classification. The method proposed in this study has been tested on Pavia University (PU) and Salinas Valley (SV) datasets, respectively. Compared with that of the other four methods, the classification performance of the method proposed in this study is significantly improved.
2023, 32(5):36-44. DOI: 10.15888/j.cnki.csa.009055
Abstract:When the artificial potential field method is employed for unmanned aerial vehicle (UAV) path planning, there are often some problems, such as unreachable targets, repeated motion trajectories, and large path lengths. The traditional artificial potential field method fails to adjust the repulsion coefficient according to the specific information of the environment, while the existing improved methods cannot take into account the planning effect and planning time while adaptively adjusting the repulsion coefficient. To solve the above problems, this study proposes a UAV path planning method based on the adaptive repulsion coefficient with the help of deep learning. Firstly, the most suitable repulsion coefficient sample set in a specific environment is found by integrating a genetic algorithm and the artificial potential field method. Secondly, a residual neural network is trained with the sample set. Finally, the repulsion coefficient adapted to the environment is calculated by the residual neural network, and then the artificial potential field method is used for path planning. Simulation experiments show that the proposed method solves the abovementioned problems in path planning with the artificial potential field method to a certain extent. It has excellent performance in planning effect and planning time and can well meet the requirements for current environment adaptation and rapid planning in UAV path planning.
2023, 32(5):45-56. DOI: 10.15888/j.cnki.csa.009110
Abstract:In the mobile edge computing (MEC) system, users’ offloading strategies will affect energy consumption and computing cost, which in turn affects the users’ benefit. However, most of the existing studies have not considered the impact of users’ offloading strategies and resource request strategies on the benefit in the random distribution of edge servers. Therefore, this study proposes a computing offloading and resource allocation strategy based on an improved double auction algorithm. Firstly, this strategy models the interaction process between users and edge servers as a Stackelberg game and proves that there is a unique Nash equilibrium point in the game. Secondly, the users’ willingness to offload different servers and the amount of computing resource requests are calculated, and then users and the optimal server are auctioned. Finally, the traversal method is employed to exchange some transactions in the previous auction for the optimal overall benefit of the system. Simulation results show that, compared with other benchmark algorithms, the proposed algorithm can improve the total benefit of system users by 33.4% in the scenario of random distribution of servers and effectively reduce system loss.
2023, 32(5):57-66. DOI: 10.15888/j.cnki.csa.009070
Abstract:Container monitoring is one of the most important factors in ensuring the proper functioning of container infrastructure. However, the dimensions of current container monitoring are single, and there is a lack of intuitive and effective methods to assist operation engineers in locating the cause of anomalies in the business container. Therefore, this study proposes a multi-dimensional container monitoring system for Kubernetes. Through the correlation analysis of monitoring metrics, it provides highly relevant metrics as the core metrics of multi-dimensional monitoring for operation engineers and thus better implements comprehensive container monitoring.
2023, 32(5):67-76. DOI: 10.15888/j.cnki.csa.009105
Abstract:For better data application and management capability of air-force stations in information construction, this study proposes a Spark-based visualization platform for flight security big data at the stations. On the basis of the data collected with the station informatization system and Internet of Things (IoT) data, a Spark computing engine is used to integrate Kafka message queues, and Hive is used for the establishment and storage of a data list library. Data pre-processing and interaction are completed on the basis of Spark RDD and Spark SQL, and the Vue framework is selected to embed ECharts components to visualize front-end data display. Finally, the design solution is implemented and applied. Compared with the current business-isolated information-system construction mode, the platform has higher abilities for data fusion and processing analysis and can better realize the value of flight security data at the stations.
2023, 32(5):77-86. DOI: 10.15888/j.cnki.csa.009103
Abstract:Live weather data should be of high uniformity, accuracy, and timeliness when all units of the meteorological department provide external services. To ensure this, this study designs a block data processing, storage, and query model for live weather data with a resolution of 1 km in China given application scenarios based on the user’s location service by the block integration method with a longitude interval of 1°. Moreover, the microservice application mode is adopted to develop a service interface of live weather data. The data service delay is improved to a minute level, and the system supports more than 200000 concurrent visits. This provides technical support for solving the problem of inconsistency in live weather data released by different applications in the same location at the same moment. As of December 2021, the system had provided live weather data to 54 meteorological service applications in China, with more than 100 million visits per month. The high concurrency, timeliness, and availability of the system have been verified.
2023, 32(5):87-96. DOI: 10.15888/j.cnki.csa.009102
Abstract:The electric submersible pump well system is an important tool for oilfield exploitation owing to its advantages of large displacement, high head, and flexible operating environment. Reducing the hazards of the faults in the electric submersible pump well system requires the fault components to be quickly and precisely located and repaired. This study proposes a knowledge graph-based fault diagnosis method for electric submersible pump wells. The improved bi-directional long short-term memory-conditional random field (BiLSTM-CRF) entity identification algorithm and the bidirectional encoder representations from transformers (BERT) relation extraction algorithm are used to extract expert knowledge from fault data and then construct a knowledge graph in the field of fault diagnosis of electric submersible pump wells; a Bayesian inference network with fault signs as initial nodes is built with the constructed knowledge graph, and the cause of the fault is inferred by utilizing historical fault data and the calculation method of decoupling conditional probabilities. The proposed method is validated by real fault diagnosis cases.
2023, 32(5):97-104. DOI: 10.15888/j.cnki.csa.009087
Abstract:The sharing of medical data among medical institutions plays an important role in realizing cross-hospital diagnosis and promoting the development of medical research. In order to solve difficult medical data sharing among medical institutions, this study proposes a cross-chain-based medical data security sharing scheme. The scheme adopts a cross-chain architecture combining relay chains, cross-chain gateways, and data chains and conducts fine-grained medical data access control based on AES and CP-ABE encryption algorithms. Moreover, searchable encryption is used to realize the safe search of medical data. In order to alleviate the computational storage overhead of blockchain, IPFS and the blockchain are combined for off-chain storage of ciphertext and on-chain storage of ciphertext addresses and keys. The security analysis and experiment prove that the scheme is feasible in medical data security sharing.
2023, 32(5):105-111. DOI: 10.15888/j.cnki.csa.009078
Abstract:The current multi-view stereo (MVS) depth estimation algorithms based on neural networks involve a large number of parameters and serious memory consumption, which is difficult to meet the needs of the current embedded platforms with low-computing power. Therefore, this study proposes an MVS depth perception network (Mobile-MVS2D) based on the MVS2D epipolar attention mechanism and MobileNetV3-Small. The network adopts the structure of encoder-decoder and uses MobileNetV3-Small network for encoding feature extraction. In addition, it adopts the epipolar attention mechanism for the coupling of scale information of different feature layers between the source image and the reference image and introduces SE-Net and jump connection to expand the decoding feature details in the decoding stage and improve the prediction accuracy. Experimental results show that the proposed model shows high accuracy in the evaluation index of depth maps in the ScanNet data set. By Combining with visual SLAM, the model can show a more accurate three-dimensional reconstruction effect and has excellent robustness. On the Jeston Xavier NX, the model only costs 0.17 s in inferring the image group with the accuracy of Float16 and the size of 640×480, and the GPU consumption is only 1 GB. Therefore, it can meet the needs of embedded platforms with low-computing power.
2023, 32(5):112-122. DOI: 10.15888/j.cnki.csa.009088
Abstract:To improve the precision of equipment defect detection of substations on the premise of only a small number of labeled samples used, this study proposes an improved defect detection algorithm based on the self-supervised model SimSiam. Unlike the original SimSiam, the improved algorithm directly utilizes non-iconic images, such as those in the dataset COCO, rather than using iconic images, like those in the dataset ImageNet, for contrastive learning and achieves performance comparable to that of any supervised methods in downstream defect detection tasks. By replacing multi-layer perception (MLP) networks with fully convolutional networks and spatial attention modules in the projection and prediction heads, the proposed algorithm preserves the spatial structure and local information of high-dimensional features. Furthermore, the output feature map is mean-pooled before similarity is calculated to obtain the eigenvector, which is then normalized to calculate the Euclidean distance and further modify the loss function of self-supervised contrastive learning. The experimental results show that the improved algorithm can make full use of non-iconic images for contrastive learning and improve the precision of equipment defect detection of substations on the premise of labeling only a small number of samples. Its mean average precision (mAP) reaches 83.84% in the detection of five types of defects, namely, broken meter dials, hanging suspended substances, nests, respirator silicone discoloration, and abnormal closure of the box door.
2023, 32(5):123-131. DOI: 10.15888/j.cnki.csa.009082
Abstract:The observation and counting of red blood cells, white blood cells, and platelets in the blood are an important basis for clinical medical diagnosis. Abnormal blood cells mean that there may be blood-related problems such as clotting abnormalities, infections, and inflammation. As artificial blood cell detection is not only labor-intensive but also prone to false detection and misses, a novel blood cell detection algorithm YOLOv5-CBF is proposed to address the above problem. On the basis of the YOLOv5 framework, the algorithm improves detection accuracy by adding a coordinate attention (CA) mechanism to the backbone network. The FPN+PAN structure in the neck network is changed to the feature fusion structure combining the idea of the bidirectional feature pyramid network (BiFPN), a cross-scale feature fusion method; in this way, the multi-scale features of the target can be effectively fused. In addition to the three-scale detection, a small target detection layer is added to improve the identification accuracy of small target platelets in the dataset. The results of a large number of experiments conducted on the dataset BCCD show that the algorithm presents an average accuracy improvement of 2.7% in the detection of the three blood cells compared to the conventional YOLOv5 algorithm, demonstrating good performance. The algorithm is highly practical for blood cell detection.
2023, 32(5):132-140. DOI: 10.15888/j.cnki.csa.009085
Abstract:To solve missing and false detection caused by different fuzzy deformations and insufficient features of target non-motor vehicles due to different heights and angles of detectors under road monitoring, this study proposes a non-motor vehicle detection model based on one-shot aggregation (VovNet) network and deformable convolution. CSPVovNet proposed by the VovNet network combined with the characteristics of the original network is used to replace the original CSPDarknet backbone network for feature extraction. This enhances the reuse of effective features and alleviates the further loss of features of small target objects caused by deep convolution. Deformable convolution is introduced into different network layers to replace the traditional convolution. Training and testing are carried out on the public data set Pascal VOC2007 and the self-built non-motor vehicle data set, respectively. The YOLOv5-C scheme is selected according to the final performance. The improved network selects EIoU_loss as the location loss. The ablation experiment shows that the final improvement improves the network performance, with the final network optimization result being 4.14 percentage points higher than the original YOLOv5s network in terms of mAP, which thus effectively alleviates missing and false detection.
2023, 32(5):141-148. DOI: 10.15888/j.cnki.csa.009061
Abstract:To effectively improve the quality of multi-image encryption and its security for data transmission, this study proposes a multi-authority multi-image encryption algorithm based on a hyperchaotic system. Specifically, bilayer cross-coupling based on the piece-wise linear chaotic map (PWLCM) is applied to L plaintext images respectively. The results are merged by exclusive-OR (XOR) to obtain a noise-like image. Then, the least significant bit embedding algorithm is used to embed the noise-like image into the (L+1)th plaintext image to obtain a semi-encrypted image. Finally, a one-dimensional cubic map is combined with a one-dimensional tent map to generate a two-dimensional cubic-tent modular map (2D-CTMM). A ciphertext image is obtained by two-step scrambling of the semi-encrypted image after it is diffused with the 2D-CTMM. The experimental results show that the proposed method, highly sensitive to plaintext and key with a large key space, can effectively resist statistical attacks and differential attacks. Moreover, the proposed algorithm enables multi-authority decryption and partial decryption by different authorized users.
2023, 32(5):149-156. DOI: 10.15888/j.cnki.csa.009064
Abstract:The trajectories of fishing vessels in the field of marine fisheries are spatiotemporal and non-stationary. Considering the problems of insufficient data extraction and low recognition accuracy in the current operation mode recognition methods for fishing vessels, an operation mode recognition model for fishing vessels, i.e., 1DCNN-SAGRU, is proposed. This model is based on the one-dimensional convolutional neural network (1DCNN) and the gated recurrent unit (GRU) network with self-attention. The model uses 1DCNN and GRU to fully extract local spatial features and temporal dependencies of the trajectory data of fishing vessels. In addition, the self-attention mechanism is introduced to strengthen the model’s ability to focus on key information. Finally, the dropout method and the RAdam optimizer are introduced to improve and optimize the model, which can prevent the overfitting of the model, speed up the convergence, and raise the output accuracy of the network. Experiments and analysis show that compared with the accuracy of other comparative models, the accuracy of this model can be improved by up to 4.4 percentage points. This indicates that the model can more accurately identify the trawl, purse seine, and gill net operations of fishing vessels, which is conducive to strengthening the regulatory capacity of fishing vessels and the protection of fishery resources.
2023, 32(5):157-163. DOI: 10.15888/j.cnki.csa.009099
Abstract:Chinese notional words are combinatorial and metaphorical in nature, and there is a lack of data sets on Chinese notional word discrimination. As a result, the understanding and discriminative capability of traditional methods for Chinese notional words are still limited in machine reading comprehension tasks. For this reason, a large-scale (600k) Chinese notional word discrimination cloze data set (CND) is constructed. In the dataset, a notional word in a sentence is replaced with a blank placeholder, and the correct answer needs to be selected from the two candidate notional words provided. A baseline model, RoBERTa-based notional word discrimination model (RoBERTa-ND), is designed to select candidate words. The model first extracts semantic information in the context using a pre-trained language model. Second, the semantics of candidate notional words are fused, and the scores of candidate words are computed by a classification task. Finally, the model’s ability to discriminate Chinese notional words is further enhanced by enhancing the model’s perception of locations and orientation information. Experiments show that the model achieves the accuracy of 90.21% on CND, beating mainstream cloze test models such as DUMA (87.59%) and GNN-QA (84.23%). This work fills the gap in the research on Chinese metaphorical semantic understanding and can develop more practical value in improving the cognitive ability of Chinese Quiz Bot. The codes of CND and RoBERTa-ND are open-source: https://github.com/2572926348/CND-Large-scale-Chinese-National-word-discrimination-dataset.
2023, 32(5):164-171. DOI: 10.15888/j.cnki.csa.009053
Abstract:Flying probe testing machines have a long detection time and low test efficiency, and their probes are easy to strike in single probe detection when detecting circuit boards. Therefore, a test path planning algorithm based on an improved particle swarm optimization algorithm is proposed. Firstly, the collision between two probes is solved by partition detection. Secondly, an improved particle swarm optimization algorithm is proposed, and a chaotic initialization formula is added to constrain and update the maximum speed of search based on the particle swarm optimization algorithm. In addition, the idea of crossover and variation of the genetic algorithm is introduced to improve some defects that the particle swarm optimization algorithm tends to fall into local optimum, which enhances the global search ability of the algorithm. The effectiveness of the proposed algorithm, particle swarm optimization algorithm, and genetic algorithm is compared and analyzed, and real machine tests are carried out. The results show that the proposed algorithm can effectively solve the collision between two probes during the tests. Compared with the other two algorithms, the improved particle swarm optimization algorithm has a stronger global search ability while reducing the number of iterations, and it can reduce the algorithm operation time by 30% and the test distance by 10%, which has a certain engineering application value.
2023, 32(5):172-179. DOI: 10.15888/j.cnki.csa.009056
Abstract:In order to detect and recognize traffic signs on the road in real time, a traffic sign recognition model based on improved YOLOv5 is proposed to solve the problems of low recognition accuracy and serious false detection and missing detection of small traffic signs under the influence of poor lighting. First, a concat operation is added to the shallow feature layer of the YOLOv5 model, and the shallow feature information is combined with the middle feature layer and then serves as a detection head, which is conducive to the recognition efficiency of small traffic signs. Secondly, a coordinate attention mechanism is added to the YOLOv5 model to improve the efficiency of feature extraction. The Chinese traffic sign dataset TT100K is expanded, and the dark light is enhanced. Finally, the improved model detection effect is verified on the preprocessed TT100K dataset. The experimental results show that the recognition efficiency of the improved model for small and dim traffic signs is greatly improved. Compared with the results of the original YOLOv5 model trained on the expanded dataset, the accuracy of the improved YOLOv5 model in this study is improved by 1.5%, reaching 93.4%. The recall rate is increased by 6.8%, reaching 92.3%. The mAP value is increased by 5.2%, reaching 96.2%.
2023, 32(5):180-187. DOI: 10.15888/j.cnki.csa.009068
Abstract:Faced with numerous online learning resources, learners often suffer from information overload and information disorientation problems. It has become a hotspot to help learners efficiently and accurately obtain suitable learning resources to improve their learning effects. Considering the deficiencies of existing approaches, such as the poor interpretability as well as the limited efficiency and accuracy of recommendation, a new recommendation approach of personalized learning resources is proposed on the basis of knowledge graphs and graph embeddings. In this approach, a knowledge graph of the online learning environment is established through a generic ontology model, and the graph embedding algorithm is applied to train the knowledge graph for optimized efficiency of graph computation in learning resource recommendation. Then, the learners’ interest in learning resources is optimized via clustering based on the learning style features of learners. Finally, the ranked recommendation results of learning resources are obtained. The experiments demonstrate that the proposed approach significantly improves the computational efficiency and the accuracy of personalized learning resource recommendations compared with existing methods in large-scale graph data scenarios.
2023, 32(5):188-195. DOI: 10.15888/j.cnki.csa.009073
Abstract:In the electronic industry, defect detection of printed circuit board (PCB) has become more and more important. Some minor or irregular damage of PCBs is closely related to visual texture information, such as dense and complex PCB cables. Feature vectors extracted from the traditional convolutional neural network are prone to lose the intermediate visual feature information such as texture features, which results in an insignificant detection effect for minor and irregular damage. To solve this problem, this study proposes a PCB damage classification model based on a Siamese deep feature fusion residual network, and the model’s backbone network is ResNet50. In the feature extraction stage, the intermediate visual features such as texture information and the high-level semantic features finally output by the neural network are fused into a 32-dimensional feature vector. The similarity between the vectors of the two features is represented by the L2 distance, which is used to judge whether the PCB is defective. Triplet loss and cross-entropy loss are applied in the training phase, and the combination of multiple loss functions improves the accuracy of the network. The validity of the model is verified by experiments, and the accuracy on the test data set reaches (95.42±0.31)%. This indicates the feasibility of the model in PCB defect classification and detection.
2023, 32(5):196-203. DOI: 10.15888/j.cnki.csa.009106
Abstract:Artificial neural networks have been developed and widely applied in computer vision and brain-like intelligence. In the past decades, research on neural networks focuses on higher accuracy rates but neglects the control of network computational costs. The human brain, as an efficient and energy-saving network, plays an important role in the development of artificial intelligence. How to emulate the connectivity properties of biological brain networks and build an ultra-low energy artificial neural network model for achieving essentially the same correct target recognition rate has become a hot research topic. To build an ultra-low artificial neural network model, this study realizes network efficiency by combining the connection properties of brain networks to change the connections of artificial neural networks. The experimental results show that combining the connectivity properties of biological brain networks to change the connections of the networks largely reduces the computational cost of the network, while the performance of the network is not significantly affected.
2023, 32(5):204-211. DOI: 10.15888/j.cnki.csa.009083
Abstract:Aiming at the problem of limb movement limitation caused by severe neuromuscular disease in patients with intact brain cognition, this study tries to enable patients to autonomously control their impaired limbs again and thus proposes an electroencephalogram (EEG) classification method of robotic arm grasping task to carry out rehabilitation training for patients with impaired limb movement. Firstly, the motor imagery EEG signals are collected by a non-invasive EEG technology, and they are then classified and identified by preprocessing, feature extraction, and convolutional neural networks with multi-scale feature fusion. Finally, the label obtained by the classification model is decoded into instructions that can be recognized by the robotic arm, so as to control the arm to fulfill specific tasks. The experimental results show that the EEG data collected from experimentally selected 15 healthy subjects in the motor imagery experiments are feasible, and the average accuracy rate reaches more than 82%, which provides a new idea for EEG classification of robotic arm grasping task.
2023, 32(5):212-219. DOI: 10.15888/j.cnki.csa.009107
Abstract:DeepLabV3+ ignores the loss of part of detail information due to the importance of features at different scales in the feature extraction stage, which results in imprecise image segmentation. In response, this study proposes an improved algorithm integrating dual-branch feature extraction and attention mechanism. The feature map extracted by the ResNet101 backbone network is used as the input feature of the attention mechanism, which solves the problems of network degradation and gradient disappearance and also captures the image details ignored by DeepLabV3+. The dual-branch feature extraction mechanism expands the feature extraction capability and refines the image edge information to optimize the uneven attention of the network to features at different scales. At the same time, the CE loss function and the Dice loss function are jointly used to reduce the influence of background by focusing on foreground samples and improve segmentation accuracy. The experimental results show that the mean intersection over union (MIoU) of the improved algorithm on the PASCAL VOC 2012 and CityScapes datasets reaches 79.92% and 68.59%, respectively. Compared with the classical algorithm and other improved algorithms based on DeepLabV3+, the proposed algorithm obtains a better segmentation effect.
2023, 32(5):220-226. DOI: 10.15888/j.cnki.csa.009104
Abstract:The standard Poisson multi-Bernoulli (PMB) filter for extended targets can hardly track spawning targets effectively. To resolve this problem, this study proposes an improved PMB tracking algorithm. The algorithm uses a random matrix method to model shapes and dimensions of extended targets and adopts a multi-hypothesis model to predict spawning targets in the filtering prediction stage and obtain multiple hypothetical components of gamma Gaussian inverse Wishart (GGIW). Finally, it updates the predicted components in the filtering update stage to estimate the motion state and expansion shapes of extended targets. Simulations show that the proposed algorithm has better tracking performance for spawning extended targets in comparison with the standard PMB filtering algorithm.
2023, 32(5):227-233. DOI: 10.15888/j.cnki.csa.009071
Abstract:The current high-resolution remote sensing image segmentation model based on deep learning has the problems of high delay and low response caused by a large number of parameters and complex calculations. Considering the problems, this study proposes a lightweight remote sensing feature segmentation method, which can better balance speed and accuracy. This method uses MobileNetV2 for rough feature extraction, constructs spatial information embedding branches to achieve fine feature extraction on different scales, and introduces dense connections between different levels to obtain dense contextual information. The decoding end designs the feature fusion optimization strategy to fuse the features of different scales layer by layer to increase the perception of fine-grained features. Meanwhile, upsampling with alternating deconvolution and bilinear interpolation is employed to reduce the image edge information loss. Finally, the cross-entropy loss is combined with the Dice loss to accelerate network convergence. Comparative experiments are carried out with several commonly used semantic segmentation methods to verify the effectiveness of the proposed method. The experimental results show that the segmentation accuracy of the proposed algorithm is 93.7%, and the MIoU is 88.01%, which can achieve effective segmentation of ground objects.
2023, 32(5):234-243. DOI: 10.15888/j.cnki.csa.009072
Abstract:Mosquitoes are the transmission media of various diseases. The monitoring of vector mosquitoes is the key to preventing mosquito-borne diseases. Traditional manual identification methods of vector mosquitoes have high costs and low efficiency. Therefore, a classification method of vector mosquitoes under deep learning is proposed, which is based on transfer learning and three ImageNet pre-training models including fine-tuning ResNet18, DenseNet121, and MobileNetV2. K-fold cross-validation is adopted under small data sets with 900 mosquito images, and Aedes aegypti, Aedes albopictus, and Culex mosquitoes are classified to evaluate model performance. The average peak accuracy reaches 95%, 97%, and 97%, respectively. Finally, 344 mosquito images are predicted by using the model retrained under the data sets with 900 mosquito images. Specifically, the lightweight model MobileNetV2 achieves the highest precision, recall, and F1 score all of 0.95. According to the final prediction accuracy of the three models, it is concluded that the lightweight model MobileNetV2 performs better under a small number of data sets. The experiment changes the previous model fine-tuning modes. The learning rate of the model classification layer is set to be 10 times that of the previous layer, and the prediction accuracy of Aedes albopictus is improved by 5%–6% compared with previous experiments, which solves the training convergence problem of a small number of data samples and further expands the applicable environment for vector mosquito recognition.
2023, 32(5):244-252. DOI: 10.15888/j.cnki.csa.009081
Abstract:As the retrieval-based question and answer (Q&A) technology becomes increasingly mature, determining how to effectively use existing models and retrieval tools to achieve overall optimization of Q&A systems is a practical problem that needs to be studied. The study proposes a three-stage and a fusion of feature and representation problem retrieval model (TSFR-RM) for constructing intelligent customer service Q&A systems. Firstly, the similarity between the text representations of users’ questions and questions in knowledge bases is calculated by deep learning methods to target the top-k candidate answer set and give the model the ability of generalized retrieval. Secondly, multi-angle semantic features are constructed for pairs of answers to users’ questions and questions in knowledge bases to perform accurate comparison calculations. Finally, a state prediction model is built to return accurate answers to question retrieval. The experimental and practical application results show that the model improves the accuracy of the intelligent customer service retrieval system for a cultural tourism institution by up to 9.3 percentage points in the performance index of precision compared with other feature- and representation-based question retrieval models.
2023, 32(5):253-261. DOI: 10.15888/j.cnki.csa.009131
Abstract:The distantly-supervised relation extraction method aims to efficiently construct a large-scale supervised corpus and apply it to the task of relation extraction. However, constructing the corpus by distant supervision brings two major problems: noise labels and long tail distribution. In this study, a novel distantly-supervised relation extraction model is proposed. Unlike the previous pipeline-based training, an external knowledge enhancement module is added in addition to the sentence encoder module. By preprocessing and coding the existing entity types and relations in the knowledge base, the external knowledge that the sentence package text does not have is provided for the model. It is conducive to alleviating the problem of insufficient information caused by insufficient long tail relation instances in the data set and improving the discrimination ability of the model to noise instances. Through a large number of experiments on the benchmark data sets NYT and GDS, the AUC value has increased by 0.9% and 5.7% respectively, compared with the mainstream optimal model, which proves the effectiveness of the external knowledge enhancement module.
2023, 32(5):262-272. DOI: 10.15888/j.cnki.csa.009109
Abstract:Crossing behavior detection is of great significance for epidemic control and social security and can reduce accidents caused by illegal crossing behavior to a certain extent. In view of the problems of poor real-time performance and the need for prior knowledge in the current crossing behavior detection task, this study applies the Faster RCNN+SlowFast spatiotemporal behavior detection algorithm to the crossing behavior detection task to split and detect the crossing behavior. In order to improve the detection accuracy and speed of the target in the spatiotemporal behavior detection algorithm, the target detection module, namely Faster RCNN is changed to lightweight YOLOv5 with high real-time performance. Then, according to the extensive in-class diversity under different perspectives of the same behavior, the residual block of the Fast branch and Slow branch is changed to AC residual block and SE residual block, respectively, so as to strengthen the network’s learning ability to key features and fine-grained features. Finally, the crossing behavior detection algorithm is designed to detect the continuity of climbing and descending states. Experimental results show that the average accuracy of the network reaches 93.5%, which shows excellent performance in crossing behavior detection.
2023, 32(5):273-282. DOI: 10.15888/j.cnki.csa.009100
Abstract:As texts generated by video descriptions are of low quality and not novel, this study proposes a codec model based on feature reinforcement and text knowledge supplementation. In the coding stage, the model enhances the fine-grained feature extraction of static objects in a video by strengthening local and global features, thus improving the resolution of similar semantics of objects. Then, it integrates visual semantics and video features into a long short-term memory (LSTM) network. In the decoding stage, to mine the hidden information that can hardly be discovered by machines in the video, the model intercepts partial video frames and detects the visual goals in them. Then, the obtained visual goals are used to extract knowledge from the external knowledge base to supplement the generation of descriptive texts and thus produce more novel and natural text descriptions. The experimental results on datasets MSVD and MSR-VTT demonstrate that the proposed method shows good performance, and the generated content can show novel implicit information to a certain extent.
2023, 32(5):283-290. DOI: 10.15888/j.cnki.csa.009054
Abstract:In the tanker loading and unloading area in a chemical plant area, preventing the generation and harm of static electricity in the tanker is an important means to avoid the combustion and explosion of the tanker. The static electricity induced by the tanker can be conducted away by the electrostatic grounding line to avoid sparkover with external substances. How to ensure that the grounding line is correctly installed during the loading and unloading process and will not be accidentally disassembled or disassembled in advance is an urgent problem to be solved in a plant area. To ensure that real-time images can be detected in real time when explosion-proof cameras are used in the explosion-proof area, this study gives due consideration to the characteristics, including different connection angles and thinning under stretching, of grounding lines and proposes a deep learning you only look once version 5 (YOLOv5) target detection algorithm by introducing the self-attention mechanism CotNet. The detection speed and detection accuracy of the proposed algorithm are compared on a self-made grounding line dataset. The experimental results show that the improved YOLOv5 algorithm, increasing the detection accuracy by 5% at the cost of a slight decrease in speed, can meet the needs of on-site grounding line detection.
2023, 32(5):291-299. DOI: 10.15888/j.cnki.csa.009066
Abstract:Semantic entailment recognition aims to detect and judge whether the semantics of two Chinese sentences are consistent and whether there is an entailment relationship. The existing methods, however, usually face the challenges of Chinese synonyms, polysemy, and difficulty in understanding long texts. To solve the above problems, this study proposes a co-driven Chinese semantic entailment recognition method based on the fusion of Transformer and sememe knowledge of HowNet. First, the internal structural semantic information of Chinese sentences is encoded at multiple levels and undergoes data-driven processing by Transformer. The external knowledge base HowNet is introduced for knowledge-driven modeling of the sememe knowledge correlations between words. Then, the interaction attention is calculated by Soft-Attention and achieves knowledge fusion with the sememe matrix. Finally, BiLSTM is used to encode the semantic information of the conceptual layer of texts and infer and judge the semantic consistency and entailment relationship. The proposed method employs the sememe knowledge of HowNet to solve the problems of polysemy and synonyms and uses the Transformer strategy to resolve the challenge of long texts. The experimental results on financial and multi-semantic interpretation pair data sets such as BQ, AFQMC, and PAWSX show that compared with lightweight models such as DSSM, MwAN, and DRCN and pre-trained models such as ERNIE, this model can effectively improve the recognition accuracy of Chinese semantic entailment (an increase of 2.19% compared with that of the DSSM model) and control the number of model parameters (16 M). In addition, it can also adapt to entailment recognition scenarios of long texts with no less than 50 words.
2023, 32(5):300-307. DOI: 10.15888/j.cnki.csa.009074
Abstract:The partial maximum satisfiability problem is an important variant of the satisfiability problem. It can handle both hard and soft constraints simultaneously and thus can model a wide range of realistic problems. Local search solvers are the mainstream method to find high-quality solutions to the partial maximum satisfiability problem, and they rely on initial data states of problem instances. Aiming at the initial solution generation process of a local search solver, namely, SATLike3.0, this study proposes an improvement strategy that gives priority to satisfy the hard constraints, and the obtained algorithm is dubbed HFCRP-F. The algorithm works on the stages of initial solution construction and initial weight configuration, including propagating unassigned variables in unsatisfied hard constraints and adding initial weights to constraints based on found solutions, so as to guide the subsequent local search process. HFCRP-F and SATLike3.0 are tested by using data sets from MaxSAT Evaluation 2018–2021. The results reveal that HFCRP-F performs much better than SATLike3.0 in processing weighted instances and shows nearly the same performance as SATLike3.0 in processing non-weighted instances.
2023, 32(5):308-315. DOI: 10.15888/j.cnki.csa.009084
Abstract:The complex time correlation and high dimension of multivariable time series lead to poor anomaly detection performance. In view of this, an unsupervised anomaly detection model of multivariable time series based on graph autoencoders (GAEs) is proposed with the adversarial training framework as the basis. First, features are converted into embedded vectors. Secondly, the divided time series and embedded vectors are converted into graph-structured data. Then, two GAEs are used to simulate the adversarial training and reconstruct data samples. Finally, the anomaly is determined according to the reconstruction error of the test data under the model training. The proposed method is compared with five baseline anomaly detection methods. The experimental results show that the proposed model achieves the highest F1 score on the test data set, and the overall performance F1 score is 28.4% higher than that of the latest anomaly detection model, namely, USAD. Therefore, it can be seen that the proposed model can effectively improve the performance of anomaly detection.
2023, 32(5):316-322. DOI: 10.15888/j.cnki.csa.009065
Abstract:Liver cancer remains the main cause of cancer-induced death in the world. Nowadays, vascular interventional therapy is the main treatment method for liver cancer, and vascular imaging plays a key role in this process, providing important reference for professional doctors. However, manual labeling of blood vessels is a complex and time-consuming task, so the automatic segmentation of liver blood vessels is of great significance for related work. In this study, an attention gating unit is introduced to improve the extraction of network information, and a new network structure, UNetR-AGM, is proposed by combining this unit with the UNetR network. The balanced filtering strategy is used for pre-processing abdominal computed tomography (CT), which not only improves the contrast between blood vessels and surrounding tissues but also completes the rough segmentation of blood vessels. To verify the effectiveness of the proposed method, this study compares UNetR-AGM with other research methods on the medical segmentation decathlon (MSD) dataset and analyzes the accuracy of the algorithm. The experimental results show that the method developed in this study is more effective than other models.
2023, 32(5):323-329. DOI: 10.15888/j.cnki.csa.009079
Abstract:In view of the defects of Hyperledger Fabric in the sorting stage, an optimization scheme of graph sorting based on the corresponding comparison graph is proposed. As the corresponding comparison has a graph merging process with correlation invariance and a short algorithm running time, a topological algorithm based on transaction importance is designed to reduce the serialization conflict caused by the default sequence sorting. The experiments and analysis show that this scheme effectively solves the serialization conflict problem of the original scheme, reduces the proportion of invalid transactions in the system, improves the transaction efficiency of the system, and saves a lot of computing and storage resources.
2023, 32(5):330-337. DOI: 10.15888/j.cnki.csa.009101
Abstract:Recently, deep learning has become more and more widely used in geology. An important topic in geological modeling is building a subsurface model according to sparse spatial observation data. Deep learning-based geological modeling has been explored through conditional generative adversarial networks, which results in realistic geological images in line with spatial measurements. However, most methods are only conditioned on spatial observations, ignoring the adjustment of geological attributes in images. This study proposes a method to adjust geological images by introducing geological attribute labels on the basis of spatial measurements. The method introduces label data representing a geological attribute category as one of the generation conditions and expands an attribute classifier to cooperate with the label to adjust the generated image, achieving more controllable images. Considering the high cost of manual labeling, this study adopts semi-supervised clustering to automatically assign labels to unlabeled data using a small amount of labeled data. In addition, clustering may produce noise labels that affect the modeling results. In response, the symmetric cross-entropy loss is used to improve the classful network to enhance the robustness of the network against noise labels. Experiments are carried out on a geological dataset in the Yellow River. Results show that the method achieves realistic geological images featuring different geological patterns and conforming to spatial observations for different attribute labels, which proves the effectiveness of the method.
2023, 32(5):338-343. DOI: 10.15888/j.cnki.csa.009077
Abstract:This study uses the UVM verification methodology to verify the serial advanced technology attachment (SATA) pathway in the self-developed high-performance secure memory system on chip (SoC) chip system. In this study, the architecture of a high-performance secure memory SoC chip and the working principle of the SATA pathway system are explained. Taking the SATA direct memory access (DMA) data transmission mode as an example, this study introduces the establishment of SATA protocol link communication and the process of data transmission. In addition, this study builds the UVM system verification platform, analyzes the SATA protocol, designs and plans the test cases at the system level, and writes the C firmware test program loaded into the system to run. In this way, programmed input/output (PIO), DMA, native command queuing (NCQ), and other data transmission pathways under SATA commands concerned at the system application level are verified. Together with the specific waveform analysis, the results show that the SATA pathway-related integration design is reasonable and meets the application requirements of the chip for the SATA data pathway, which verifies the SATA pathway for the high-performance secure memory SoC chip system.