Abstract:Typically, the options of multiple choice Machine Reading Comprehension (MRC) are not directly extracted from the given document. Thus the answers need to be summarized or rewritten or even inferred from document or from the world’s knowledge. Most existing models adopt attention mechanism to generate an interactive representation of document, question, and option. However, these models are limited by only using the given document rather than common knowledge, leading to poor result when dealing with questions requiring external knowledge assistance reasoning. To address questions requiring external knowledge assistance reasoning, we propose a novel neural model by integrating external commonsense knowledge to assist multi-step reasoning. Our model first interacts information among document, question, options, and related external knowledge by attention mechanism and then predicts answer by multi-step reasoning through the interaction results. The experimental results on the SemEval-2018 MCScript corpus show that the proposed model improves the accuracy of question answering requiring common knowledge reasoning.

Abstract:A new hybrid evolutionary algorithm is proposed to solve the parallel machine scheduling problems with step-piece deteriorating processing time. The goal is minimizing the total completion time. The algorithm uses opposing strategy and Smallest Rate First (SRF) rule to generate the initial population to improve its quality, and the algorithm considers the population diversity to accelerate the convergence of the algorithm, which improves the calculation efficiency of the algorithm. At the same time, a variable neighborhood search algorithm with 3-opt perturbation operator is added to improve the quality of the results obtained by the genetic algorithm. By simulating the experiments of different scale examples, the results are improved compared with the traditional GA and VNS algorithms.

Abstract:To determine the weight of the combined forecasting model is very important to improve the accuracy of the model. Applying the regularization and cross-validation to the combined forecasting model based on the least squares method is for studying whether the regularization and cross-validation can improve the prediction effect of the combined forecasting model. It is carried out by adding the L₁ and L₂ norm regularization terms to the optimization solution of the combined model and using leave-one-out-cross-validation in the data set. The result shows that both the L₁ and L₂ norm regularization can improve prediction accuracy of the combined model to a certain degree. Moreover, the L₁ norm regularization is better than the L₂ norm regularization for the combined forecasting model, and the more single forecasting models participating in the combined forecasting, the better the regularization improvement effect. In addition, there is a positive correlation between the cross-validation improvement effect and amount of experimental data given.

Abstract:Aiming at the problem of error in customer flow statistics caused by fast moving and shielding of the target when entering the door, this study designs the strategies of target detection, tracking, and behavior judgment of entering the door, and puts forward the method of customer flow statistics in catering industry based on deep learning. The YOLOv3-tiny model is trained by multi-data set, and the accurate detection of small target is realized. The target tracking algorithm of multi-channel feature fusion is designed to achieve the stable tracking in the case of fast target movement. In this study, we design a method to judge the entry behavior of the target through overlapping rate, and realize the accurate statistics of the entrance passenger flow. The experimental results show that the average accuracy rate of passenger flow statistics is 93.5%.

Abstract:Rebar is a widely used building material, if its size and surface defects cannot be found in time in the rolling process, it will produce a large number of waste products and bring losses to the enterprise. In this study, we design a rebar surface defect detection method based on machine vision. Firstly, the skew rebar in the image is corrected by affine transformation, and then the front and side images of rebar are distinguished based on Hough transform to detect the straight line position of longitudinal rib edge. Finally, defect detection is carried out for front and side images to quickly and accurately judge whether there are defects on the surface. Experiments show that the design method has sound stability and practicality. It can effectively solve the problems of low efficiency and high false detection rate in the process of manual detection.

Abstract:In order to optimize the method of real-time and high-precision detection of drivers' safe driving supervision, based on the classic deep learning neural network-YOLOv3-tiny-in object detection, this study successfully uses the channel pruning technology to achieve model compression in the object detection task, and reduces the calculated total amount and parameters of the improved neural network under the condition of constant accuracy. Based on NVIDIA’s inference platform TensorRT, model level fusion and half-precision acceleration are performed, and the accelerated model is deployed. The experimental results show that the speed of inference of the acceleration model is about 2 times that of the original model, the parameter volume is reduced by half, and the accuracy is not lost, which realizes the purpose of real-time detection under high precision.

Abstract:Neurological diseases such as stroke, spinal cord injury, and degenerative joint diseases of the lower extremities can lead to limb dysfunction. It requires targeted and repetitive training. This study designs a wheelchair-type lower limb rehabilitation robot that can switch between sitting, standing, and lying modes. The wheelchair-type lower limb rehabilitation robot, in the active walking training mode of the robot standing posture, solves the forward speed of the human body to map the speed of the wheelchair-type chassis motor, and realizes the coordinated control of the gait training of the lower limb dysfunction patient and the robot wheelchair-type vehicle body. The experimental results show that the speed curve of the lower extremity exoskeleton walking and the speed curve of the wheelchair-type car body are basically the same, realizing the coordinated control of human and robot.

Abstract:The technology of Docker based on the lightweight technology was summarized. The Version 1.13+ of Docker was transplanted and integrated into Fedora 28 system based on Loongson. The test image was made to analyze the performance of the new version of Docker scheme, the trend of performance and the bottlenecks of container’s function under different number of containers were analyzed. By comparing the performance in containers of Loongson single-way, double-way, four-way servers and a machine with AMD Ryzen 5 (2400 GB), the stability of the new version of Docker scheme on Loongson was affirmed. The difference of performance between Loongson 3A3000 and Ryzen 5 (2400 GB) under the near basic frequency was analyzed. The prospect of industry about domestic CPU chip is expected.

Abstract:Mesoscale eddy which is of great significance to human activities and marine science is a special mesoscale phenomenon in the ocean. The detection of mesoscale eddies in marine physics usually relies on parameters predefined and adjusted by experts or scanning and judging all ocean data point-by-point. These methods cannot guarantee a satisfied accuracy rate and always take a long time. In addition, the spatio-temporal statistics of mesoscale eddies are complicated, which cannot display relevant information well, and the collation and analysis work is huge. This study proposes an oceanic mesoscale eddy detection algorithm based on deep learning target detection, and it achieves high recognition precision and recall. The proposed algorithm avoids the influence of threshold selection in the methods on mesoscale eddy detection, and greatly improves the detection speed. Meanwhile, we design a visualization system which provides mesoscale eddy space-time features and ocean information. The system can meet the need for insight, description, and correlation analysis of the statistical information, feature distribution, and attribute associations of the mesoscale eddy.

Abstract:Distributed cluster environment makes real-time data computation more complex, and the correctness of stream large data processing system is difficult to guarantee. The existing large data benchmarking framework can test the performance of stream large data processing system, but there are many shortcomings such as simple application scenario design and inadequate evaluation index. To address this challenge, this study constructs a stream large data benchmarking framework for stock trading scenarios, generates high-frequency stock trading data through a flow-based data generator, and tests the performance of the system in high-speed scenarios in terms of delay, throughput, GC time, CPU resources, and so on. At the same time, the scalability of large data streaming system is verified by horizontal test. In this study, Apache Spark Streaming is used as the test system to test. The experimental results show that the performance degradation problems such as delay increase and GC time increase occur in high-speed scenarios because of the increase of input rate and parallelism of the system.

Abstract:The existing network management software is mostly developed in C++, and the interdependence among services is relatively serious, while the operation and maintenance management software mostly based on Java development. In order to realize unified management of network management services and operation and maintenance services, this study proposes a solution for developing micro-service bases based on ServiceComb micro-service architecture to manage multi-programming language application services. The system uses the service base to realize unified access management of different language services. It is proved by practice that the system can be compatible with services in two different programming languages to meet the monitoring and management of different programming language services.

Abstract:We design and develop an unattended remote monitoring system combined with software and hardware to ensure the normal operation of radar and the timely transmission of products. Use ZigBee communication mode to realize monitoring network communication, and use LINQ technology, entity framework technology, and multi-thread programming technology to compile data acquisition, processing and controlling software. The system was used in business, and the purpose of intelligent monitoring, real-time early warning, and timely processing are achieved.

Abstract:With the rapid development of outsourcing services, cloud computing has attracted an increasing number of individuals and enterprises to enjoy the services from outsourcing service providers. Moreover, fog computing further extends cloud computing to the edge of the network. In fog computing, because the end user is usually resource-constrained, the outsourcing computation tasks can be outsourced to the fog nodes. However, the mutual distrust between users and fog nodes may impede the fair payment of outsourcing services. Nevertheless, most existing solutions adopt the traditional payment mechanism, which needs a trusted authority such as a bank. In this study, in order to realize fair payment of outsourcing services, we introduce a new fair payment framework based on Blockchain in fog computing to directly transfer rewards by smart contract. Meanwhile, we present a construction to guarantee that if there is a malicious user, the honest one can get compensation. Finally, our security analysis indicates that the proposed protocol achieves correctness and fairness, and performance analysis shows that the experimental consumption is acceptable.

Abstract:The collection of atmospheric data plays a key role in the control of small unmanned rotorcraft flight. An atmospheric data acquisition system for the acquisition device of a small unmanned rotorcraft was designed by a digital MEMS sensor and TMS320F28335 chip based on the flight control needs of small unmanned rotorcraft. The space velocity required to control the flight of the unmanned rotorcraft is calculated in real time by collecting data with the sensor, and some algorithms such as linear interpolation are used to improve the calculation performance. The system can meet the needs of small unmanned rotorcraft because of its small size, low power consumption, and strong anti-interference ability, which has made some theoretical work for the practical application of small unmanned rotorcraft data acquisition system.

Abstract:This study designs a volume measurement system based on laser scanner for the logistics industry to measure the volume of goods. The system consists of two-dimensional LiDAR, servo motor, motor control unit, data acquisition and processing unit, and other mechanical devices. The data acquisition and processing unit uses discrete integral operations to process 3D point cloud information and calculate the volume of the object. The application of multimedia timers enables data acquisition, data processing, and command response to work in harmony. The results show that the system’s measurement error can be controlled below 5%, scanning at a speed of about 0.5 m/s and at a position 5 to 6 meters from the plane of the object. The system has strong applicability, which can measure the volume of all objects in the scanning area and has broad application prospects.

Abstract:In the future, the main development mode of pig breeding industry is information and intelligence. In order to monitor the behavior of pigs intelligently, so as to monitor the health and growth of pigs, this paper presents a system technology for contactless identification and monitoring pig behavior based on machine vision. The system collects pig behavior sequence images by CCD camera, then extracts the depth features of those images using convolution neural network. After that, the feature fusion method is used to fuse the depth features of the behavior sequence images. Finally, the pig’s behavior activities are identified according to the fusion depth feature. The system realizes the high-precision identification of pig’s motion behavior, claudication behavior, volt behavior, breathing behavior, eating behavior, and excretion behavior under natural scenes. The accuracy rates of recognizing all kinds of behavior are more than 94%, which are higher than the state-of-the-art methods.

Abstract:Aiming at the problems of frequent start and stop in acceleration and deceleration control method with zero first speed and uneven acceleration transition of end-effector in interpolation process, a forward-looking interpolation algorithm for smooth transition of multi-trajectory segments based on asymmetric S-shape acceleration and deceleration control is proposed. The algorithm uses the arc model to smooth the transition at the connecting corner between adjacent track segments. Given the coordinates of the connecting point of the path and the radius of the transition arc, the optimal speed of the connecting arc is planned. A new flexible acceleration and deceleration control algorithm is used to solve the normalization factor in the interpolation algorithm. The algorithm is formed by fitting the cosine acceleration and deceleration curve on the linear acceleration and deceleration curve, which reduces the computation of cosine acceleration and deceleration algorithm, and ensures the stability of acceleration control. The experimental results show that the algorithm can realize the smooth transition at the joint of multiple track segments, ensure the smoothness and continuity of motion speed, and effectively improve the operation efficiency of the terminal actuator.

Abstract:In order to solve the problem of setting cut-off distance and selecting clustering center in Density Peak Clustering algorithm (DPC), a new self-adjusting step-size fruit fly optimization algorithm is used to calculate the cut-off distance and the important parameters in density peak clustering, an adaptive method for selecting clustering centers is designed. In the cut-off distance calculation process, the search step-size is dynamically adjusted according to the rate of change of the difference between the optimal concentration and the worst concentration in each step of the iterative process, and its optimization efficiency and accuracy are better than the existing improved fruit fly algorithm. In the selection process of clustering center, the clustering center is selected adaptively according to the distribution of the product of local density and distance. The computational accuracy and efficiency of the proposed algorithm are both better than the existing improved DPC algorithm, and it can realize data clustering completely adaptively.

Abstract:The service volume of online customer service system is affected by many factors. In order to improve the prediction accuracy of service volume, an improved algorithm IDMPSADE is proposed on the basis of self-adaptive differential evolution algorithm DMPSADE with discrete mutation control parameters. By combining IDMPSADE with Long-Short Term Memory network (LSTM), an IDMPSADE-LSTM prediction model of service volume is established. IDMPSADE chooses the reverse guidance of the parent population whose child population’s performance on test functions is not as good as it, which can escape from the local optimum and improve the capability of searching the optimal solution within defined space. LSTM’s parameters, such as number of neurons, epochs, learning rate, and batch-size, are set by experience and have larger randomness, and IDMPSADE could be helpful to optimize these parameters. IDMPSADE-LSTM prediction model uses temperature and precipitation as influencing factors and combines with the temporal characteristics of service volume to predict the service volume. The experimental results show that the proposed IDMPSADE-LSTM prediction model is more accurate compared with general neural networks and SARIMA-SVM hybrid prediction model.

Abstract:By introducing an efficient statistical scheme for WAF traffic analysis processing which provides a mechanism for each processing Flow to dynamically adjust the processing capability, realizes dynamic adjustment processing capability, and realizes coordination adjustment processing capability and the synchronization mechanism between the various processing. On the other hand, in the traditional topology, drainage is started by the virtual network of the host machine or started by a stream serial thread with SDN. By introducing the high concurrent traction scheme in this study, the SDN traction Flow, strategy downward separation, and the concurrent protection of WAF are realized. Experimental results show that the scheme improves the accuracy of WAF protection and improves the WAF throughput.

Abstract:In this study, we proposed an improved method of infrared pedestrian detection and overlap rate, which could be used for positioning and abnormal alarm, especially in the security industry. The method is consisted of three steps: 1) infrared pedestrian detection algorithm; 2) classification algorithm; 3) overlap rate algorithm and the logic of alarm. Infrared sensors could collect high quality image data at night, and overcome environmental interference as much as possible. Pedestrian detection was designed by YOLOv3 algorithm and Multi-Layer Perception (MLP) based on Histogram of Oriented Gradient (HOG) features. The abnormal alarms were proposed by calculating overlap rate between pedestrian detection bound and ground truth bound, and then making logical judgment. The experiments evidenced the benefits of proposed approach, which could effectively improve pedestrian detection performance and abnormal alarm accuracy (over 91%).

Abstract:Aiming at solving the problem that the traditional ID3 algorithm is complicated and there exists redundancy information, this study proposes an improved algorithm—attribute reduct simplified ID3 algorithm based on rough set. This algorithm uses the properties of attribute reduct in rough set to delete the redundant knowledge, and makes the classification system more concise with the same classification ability. At the same time, it simplifies the entropy formula with Taylor formula to make the calculation easier. And then this study applies the improved algorithm to the example, and uses the massive data in the related database to program in order to do simulation experiments. Finally, the simulation results proved the correctness and feasibility of the proposed algorithm. It can not only reduce the information duplication, reduce the redundancy rules, but also ensure the accuracy of the algorithm. At the same time, it provides a certain reference value for the better application of ID3 algorithm to real life examples.

Abstract:In this study, we mainly aim at the problem that Wi-Fi provisioning of the smart goods in home always takes too much time but hard to troubleshoot. Based on detailed study of “Soft AP” and “Simple Config” which are most commonly used for WiFi provisioning of smart goods, we proposed a new research method in which packet capture is involved among the cell phones, smart goods, and corresponding servers to locate the issues. By applying this method to research time-effectiveness of WiFi provisioning for smart water heater and smart air conditioner, we successfully located the key segments and factors that impact the time-effectiveness. The optimization measures based on the research result have reduced the time consumed on WiFi provisioning a lot, and improved the user experience greatly.

Abstract:The problem of air environment has become the focus of attention. Apart from the exhaust emissions from factories, the popularity of private cars has led to worrisome air conditions. Related government agencis have also begun to strengthen the control of air environment, and put forward relevant policies for grid monitoring of environmental quality. In this context, many micro-monitoring instruments have emerged into the market, but due to the inadequate accuracy of internal sensors, there is a problem of data deviation. In order to solve this problem, this study uses the Long Short-Term Memory (LSTM) model of neural network technology and semi-supervised learning method to improve the accuracy of monitoring data. By comparing with other models, this method achieves a sound effect.

Abstract:Natural gas load forecasting is especially important for gas-operated enterprises. It is extremely important to ensure the gas consumption of natural gas pipeline network. The traditional natural gas prediction model has low prediction accuracy and low generalization of the model, so that accurate load prediction cannot be performed. In order to overcome these defects, a natural gas load forecasting model based on wavelet neural network optimized by genetic algorithm is proposed. The genetic algorithm is used to optimize the parameters of wavelet neural network threshold and network connection weight to establish the best prediction model. The validity, feasibility, and accuracy of the prediction model are verified by the historical gate data provided by the enterprise. The simulation results show that the wavelet neural network using genetic algorithm to optimize the network parameters improves the prediction accuracy of the model and has sound engineering application value.

Abstract:Image super-resolution reconstruction has a wide range of applications, such as security systems, small object detection, and medical imaging. This study proposes a dual stream feedback network to improve the performance of image super-resolution reconstruction. In the dual-stream network, one path adapts a deep residual dense network to learn the high-frequency information of the reconstructed image, and the other path directly samples the input image to the desired resolution through a sub-pixel convolution layer. Then, the feature maps obtained from the two paths are fused to adaptively selecting the required information. Finally, using a feedback convolutional layer for locally loop training to obtain a large receptive field. By training on the dataset DIV2K, the experimental results show the effectiveness and superiority of the proposed method.

Abstract:It is needed to be adapted to the actual engineering requirements and the classification of the fine-grained data when we collect and annotate data. However, It is difficult to maintain complete independent and identical distribution between the samples. The non-i.i.d data seriously reduce the training’s robustness of deep neural network model and the generalization performance of specific tasks. In order to overcome the shortcomings, this study proposes an improved algorithm of batch normalization, which normalizes a fix reference batch to calculate its mean and variance when the model training started, and then, the statistics of the reference batch is used to update other batches. Experimental results show that the proposed algorithm can accelerate the training convergence speed of the neural network model, meanwhile, the classification error is reduced by 0.3% compared with the BN algorithm. On the other hand, the robustness of neural network model and the generalization performance of some detection frameworks like object detection or instance segmentation are also improved effectively.

Abstract:In the existing depth estimation algorithm, when the depth estimation is performed on the image of the light field sequence, the matching effect is poor and the robustness is low when the image brightness changes and in the weak texture region. Aiming to solve these problems, this study proposes an adaptive weight block matching algorithm based on CIELab color space. Since the color difference matching in color image RGB color space has many influencing factors, the algorithm converts to CIELab space for color similarity matching to calculate the weight value, and then combines the gradient and distance to calculate the matching image and the matching block in the image to be matched to obtain the comprehensive weight value. Finally, according to the linear characteristics of the Epipolar Plane Image (EPI), the matching image and the image block to be matched in the image sequence are matched and calculated, and the depth map is obtained. After simulation, the proposed algorithm can estimate the depth information of the scene better, and the accuracy is greatly improved. It is obviously superior to the previous depth estimation algorithm and can be widely used.

Abstract:IEEE 802.11ac technology has become the main technology of the next generation LAN, which has greatly improved the transmission rate by increasing the types of multiple input and multiple output technologies, channel width, coding and modulation strategies. However, the increased types also lead to many problems such as excessive search space when the rate is selected, long search time, and high computational complexity. And the existing rate adaptive algorithm for IEEE802.11a/b/g/n cannot solve this problem. To settle this problem, we propose a hybrid rate adaptive algorithm called VhRa in high-speed wireless LAN mode. The algorithm utilizes the MIMO mode, the optimal setting of the channel width and the monotonic relationship between the RSSI to perform feature extraction, and selects the mode based on the zigzag detection by using the dichotomy in the MCS selection, thereby reducing the space for algorithm search and improving the search efficiency. Thus, the algorithm further increases the transmission throughput. The experiment results show that VhRa offers search efficiency up to 42% over RRAA, up to 20% over Minstrel-HT. At the same time, the throughput analysis of the algorithm is carried out in different scenarios. Compared with RRAA, Minstrel-HT, and Samplelite in mobile environment, VhRa offers throughput gain up to 90% over RRAA, up to 10% over Minstrel-HT, and up to 34% over Samplelite.

Abstract:Aiming at the shortcomings of the current pedestrian re-identification technology, this paper presents a pedestrian re-identification method based on Siamese network. First, Dropout algorithm is used to improve the performance of Convolutional Neural Network (CNN), which can reduce the incidence of the fitting problem. By integration of classification and inspection in the CNN, Siamese network is constructed to improve the efficiency and accuracy of image recognition. Finally, Markov distance for metric learning algorithm is used as the evaluation index of image matching similarity. Experiments are conducted on the Market-1501, and the experimental results show that this method is effective in terms of improving the efficiency and accuracy of pedestrian re-identification algorithm.

Abstract:The recognition of stop point in trajectory is the key step to transform spatial trajectory into semantic trajectory. At present, the recognition method of the stop point lacks the consideration of the time continuity of the record point, which leads to the lack of time information of the identified stop point. At the same time, in the scenario of missing track points, the information of stop point cannot be accurately identified. In order to solve these problems, this study proposes a velocity based spatio-temporal clustering method. Firstly, the real missing sub trajectories are determined by the spatio-temporal characteristics of the missing trajectories, and then the missing trajectories are interpolated according to the average velocity of the missing trajectories. In the experiment, GeoLife trajectory data is used to verify the proposed method. The results show that the algorithm can effectively identify the user's stop point, and has relatively sound robustness to the interference in the trajectory.

Abstract:In view of various situations caused by traffic jam, video analysis is used to realize real-time and efficient detection of vehicle queue length, so as to obtain more traffic information and improve traffic conditions. In this study, the improved FAST algorithm is obtained by combining the traditional FAST corner detection method with the process of motion detection. By using the improved FAST corner feature analysis technology, not only can the corner feature graph representing the presence of vehicles on the current traffic road be extracted, but also can the motion state of corner position be obtained. After the pretreatment of video under traffic monitoring, the static corner point features in a single lane form vehicle queuing, and PCA processing is carried out to obtain one-dimensional vector. Finally, morphological processing is carried out to detect the queue length of vehicles in a single lane. The experimental results show that the detection accuracy of this method is 98% on average, which can be applied to the actual scene.

Abstract:In a multi-device environment, the differences of applications on the user interfaces of different devices lead to duplication and difficulty in the interface designing work. With the application interface pattern, developers can get rid of the development method which generates user interface using tedious underlying UI controls, and focus on the macro-interactive solutions, thus providing a possible solution to multi-device interface generation problems. Based on PLML, a device-independent interface pattern description language SPLML is designed to represent the information of the pattern-based interface elements, and the user interface pattern generation framework UIPF on different platforms is used to support the automatic generation of the interface. The specific case illustrates the feasibility and effectiveness of the program.

Abstract:Accurate financial risk assessment is an effective measure to prevent and resolve corporate financial crisis. In this study, a method of enterprise financial risk assessment based on grey system theory is proposed. The method uses the grey system theory to construct the financial evaluation model, and generates the evaluation index system from the two aspects of enterprise financial indicators and non-financial indicators. The particle swarm is improved based on the chaotic model, and the improved particle swarm optimization algorithm is used to optimize the weight of the gray system. Finally, an experimental analysis is carried out with a real estate enterprise data. The results show that the method can effectively complete the financial risk assessment of enterprises and has high evaluation accuracy.

Abstract:In order to quickly and accurately obtain the global optimal solution for task allocation of distributed agile software development teams, the study proposes a method based on capability matching. This method builds the utility function of capability matching on the basis of the subtask capability demand degree vector and the team ability vector. We solve the utility matrix, and the optimal allocation scheme is obtained when the global utility value is the largest. The simulation results show that the method could effectively obtain a task allocation scheme with better capability matching.

Abstract:At present, traditional algorithms in IT recruitment information classification have long-distance dependence, and cannot highlight the impact of IT job keywords on text classification features. In this study, a multi-layer text classification model combining two-way long-term and short-term memory network BiLSTM and attention mechanism is applied to the classification of recruitment information. The model includes the one-hot word vector input layer, BiLSTM layer, attention mechanism layer, and output layer. One-hot layer builds a recruitment dictionary, which saves a lot of training word vector time; the BiLSTM layer can obtain more semantic information of different distances in the context; and the attention mechanism layer transforms the weights of the data encoded by BiLSTM enhancing the serialization learning task. The results show that the classification accuracy of IT recruitment information based on this model reaches 93.36%, which is about 2% higher than other models. The model analyzes the requirements of different positions on the ability of the employed in a more targeted manner, and realizes the classification of recruitment information in different positions, which is of great significance to the employment guidance of college graduates.

Abstract:In this study, the course recommendation prediction model based on k-NN algorithm has been built. Due to the original sample data of the local imbalance and data overlapped, the prediction score of the prediction model is not ideal without any parameter adjustment and data optimization. Aiming at the above problems, this study designed a set of parameter optimization scheme and sample data discretization algorithm of the prediction mode, including the best k value selection algorithm, distance formula optimization, and data discretization algorithm design. In the study, the design of the “data discretization algorithm” drives kd tree classification feature space order sorted by the weight of the characteristic vector that we expect, this algorithm plays a positive role in improving model prediction score. Therefore, all of that increases the grade of the model from 0.67 to 0.85, and the accuracy of prediction results is increased by 27 percentage points, and students' satisfaction with course recommendation is significantly improved.

Abstract:The detection of key points of clothing plays an important role in the classification, recommendation, and retrieval of clothing. However, there are a large number of clothing pictures with deformation and complex background in the clothing database, which leads to the poor recognition rate of the existing clothing classification model and the effect of clothing recommendation and retrieval. For this reason, this study proposes a model called Cascaded Stacked Pyramid Network (CSPN) which combines the target detection method with the regression method. First, the costume target area is identified by the Faster R-CNN, and then the Cascaded Pyramid Network (CPN) is constructed based on the multi-level feature map generated by ResNet-101 structure. This model integrates the multi-scale and different-layer clothing image feature, and solves low image recognition accuracy about clothing key points of the deformation and complex background image. Experimental results show that the CSPN model has higher recognition rate on the key points of clothing than the other three models in the DeepFashion dataset.

Abstract:The classification of forest types plays an important role in the management of forest ecosystems. Because of the large number of bands in hyperspectral imagery, the traditional methods of dimensionality reduction include features selection or feature extraction, affect the accuracy of forest type identification to a certain extent. The Deep Belief Network (DBN) is a semi-supervised learning method that uses all bands of hyperspectral image as input to avoid dimensionality reduction. Forest type identification of 8 townships in the west of Dehua County in Quanzhou was carried out. At the beginning, the classification of forest types in hyperspectral imagery was realized by Python language, according to HJ/1A hyperspectral image and forest management data. In addition, the influence of network depth and number of hidden layer units on overall accuracy and Kappa coefficient was discussed. The experimental results show that the network with 3 layers and 256 nodes is the optimal structure for forest type identification. The overall accuracy is 85.8% and the coefficient is 0.785, which is better than the classification result of support vector machine.

Abstract:Computer vision is an important branch of machine learning at present, which requests much higher instantaneity and accuracy as the driverless and SI-Drive development. To optimize the current methods, the Faster Region-based Convolutional Neural Network (Faster R-CNN) is upgraded by adding SENet to it in this study. The upgraded Faster R-CNN model is applied in pedestrian detection. The new model does not only bring higher accuracy but also accomplish a better detection rate. To verify the new method, an examine was done in INRIA set and our set. The result shows that the upgraded model has a better detection performance on both accuracy and rate which can meet the related specifications of real-time pedestrian detection basically. Finally, the method was tested in the NVIDIA GTX1080Ti GPU. The results show that the mAP of upgraded model can achieve up to 92.7%, while the detection rate is up to 13.79 f/s under a relatively plain experimental condition. On the whole, the new model performs better than the traditional Faster R-CNN model.