Online FirstList of IssuesDocument Types

    Volume51,Issue11,2023
    Issue Cover

      PAPERS

    • Decoupled Estimation Method for Position and Velocity of Moving Target Based on Passive Radar System

      WANG Ding, YIN Jie-xin, WANG Ye-lu, XU Wen-yan
      Vol. 51, Issue 11, Pages: 3011-3023(2023) DOI: 10.12263/DZXB.20230759
      Abstract:In order to locate radio silent target on the Earth surface, a novel positioning method for moving target is proposed based on passive radar system. Unlike most existing positioning methods, this method takes into account the quadratic constraints that the target position vector and velocity vector need to satisfy, and can achieve decoupled estimation of the target position and velocity. Firstly, the nonlinear observation equations based on the passive radar system are transformed into pseudo-linear observation equations. Then, the asymptotic statistical properties of the errors in pseudo-linear equations are derived by applying the first-order error analysis. Subsequently, an optimization criterion is constructed for the joint estimation of the target position and velocity under the two quadratic equality constraints. For the purpose of obtaining the global optimal solution of the target position and velocity, a decoupled optimization algorithm is developed based on the Lagrange multiplier approach. This optimization algorithm requires only the iteration of the target position parameters, and the target velocity parameters can be obtained in a closed form, thus reducing the influence of initial values and the risk of local convergence. Furthermore, the Cramér-Rao bound (CRB) for moving target localization based on passive radar system is deduced under the two quadratic equality constraints, and the performance gain resulting from the equality constraints is quantified. The new estimator is also proved to be asymptotically statistically efficient by using the first-order error analysis as well as the Lagrange multiplier approach. Finally, simulation results verify the advantages of the proposed positioning method.  
      Keywords:passive radar;moving target;Lagrange multiplier approach;theoretical performance analysis;Cramer-Rao Bound;decoupled parameter estimation  
      110
      |
      88
      |
      1
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • HAN Xu-cang, CHEN Bo-tao, CAO Wei-qiong, CHEN Hua, LI Hao-yuan
      Vol. 51, Issue 11, Pages: 3024-3032(2023) DOI: 10.12263/DZXB.20230419
      Abstract:Masking is widely used in secure implementations of modular exponentiation, and its ability of side-channel resilient has been well-demonstrated. During the modular multiplication in modular exponentiation, we discovered that there are several fetch operations, and variations in the power consumption, which revealed the address of the operands, and then proposed a template attack based on linear discriminant analysis aiming at this vulnerability. In contrast to operand-based leakage, fetch-based leakage is not affected by mask and thus can be effective in attacking masking-based modular exponentiation. In our analysis, we extended testing vector leakage detection to the extraction of leaked features, which reduced the influence of irrelevant points. Second, linear discriminant analysis was utilized to trace classification and reduced the dimensionality of traces, which improved the ability of trace separability. Finally, an attack was conducted on a hardware implementation of masking-based modular exponentiation. Results showed that fetch-based leakage was distributed in the entire modular multiplication operation, and the correct ratio of modulo multiplication identification is up to 99.98%.  
      Keywords:exponential masking;masked implementation of modular exponentiation;template attack;linear discriminant analysis;testing vector leakage detection;operand loading  
      60
      |
      71
      |
      0
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Autoimmune Dynamic Attack Generation Method Based on Reinforcement Learning

      LI Teng, TANG Zhi-liang, MA Zhuo, MA Jian-feng
      Vol. 51, Issue 11, Pages: 3033-3041(2023) DOI: 10.12263/DZXB.20230369
      Abstract:The approach of launching network attacks through optimal pathways has become a significant factor affecting the internal network security of various enterprises and organizations. Existing methods for exploring optimal attack pathways within internal networks mostly rely on attack graphs and often neglect the relationship between attack costs and benefits. Methods that utilize the Q-learning algorithm to address attack pathways suffer from low utilization of network vulnerability information. To address these issues, this paper draws inspiration from the biological immune system and proposes a reinforcement learning-based dynamic self-immune attack generation method. This method simulates network attacks by intruders on an internal network, efficiently uncovering vulnerabilities within the internal network, thereby achieving self-immune defense. The proposed approach first acquires and processes internal network information, attaches weights to directed edges in the attack graph, and then employs an improved Q-learning algorithm to discover optimal attack pathways. It successfully identifies all optimal attack pathways, providing attack graphs and an analysis of host vulnerabilities within these pathways. Theoretical analysis and experimental results demonstrate that this method not only efficiently and accurately identifies optimal attack pathways but also resolves issues such as ring loops and multiple optimal attack pathways. By making full use of internal network vulnerabilities, it enhances self-immune security defenses.  
      Keywords:optimal attack path;reinforcement learning;attack graph;path planning;intranet security  
      53
      |
      82
      |
      0
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • An Fusion SLAM Method for LiDAR Visual and IMU Based on Factor Map Elimination Optimization

      YUAN Guo-shuai, QI Yong-sheng, LIU Li-qiang, SU Jian-qiang, ZHANG Li-jie
      Vol. 51, Issue 11, Pages: 3042-3052(2023) DOI: 10.12263/DZXB.20230209
      Abstract:Addressing the limitations of single-sensor SLAM (Simultaneous Localization And Mapping) techniques, degraded perception, and poor reliability in complex environments, this paper proposes a multi-factor graph fusion SLAM algorithm with IMU as the dominant system (ID-MFG-SLAM). Firstly, the utilization of a multi-factor graph model, with the IMU (Inertial Measurement Unit) as the primary system and visual and LIDAR sensors as secondary systems. This novel structure incorporates observation factors from the secondary systems to constrain IMU biases and integrates IMU odometry factors for motion prediction and fusion. To reduce the optimization cost after fusion, a sliding window mechanism is introduced for historical state information backtracking. Additionally, a QR decomposition elimination method based on Householder transformation is employed to convert the factor graph into a Bayesian network, simplifying the graph's structure and improving computational efficiency. Furthermore, an adaptive interpolation algorithm between quaternion spherical linear interpolation and linear interpolation is introduced. This algorithm projects LIDAR point clouds onto a unit sphere, enabling depth estimation of visual feature points. The experimental results show that compared to other classic algorithms, this method can achieve absolute trajectory errors of about 0.68 m and 0.24 m in complex large and small scenes, respectively, with higher accuracy and reliability.  
      Keywords:simultaneous localization and mapping;multi-sensor fusion;complex scene;laser radar;IMU odometer;factor graph optimization  
      66
      |
      83
      |
      4
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Noise Filtering and Feature Enhancement Based Graph Neural Network Method for Fraud Detection

      LI Kang-he, HUANG Zhen-hua
      Vol. 51, Issue 11, Pages: 3053-3060(2023) DOI: 10.12263/DZXB.20230489
      Abstract:Existing graph neural network (GNN)-based fraud detection methods have at least three shortcomings: (1) They do not adequately consider the problem of imbalanced distribution of sample labels. (2) They do not take into account the problem that fraudsters deliberately create noise to interfere with fraud detection in order to avoid detection by detectors. (3) They fail to consider the limitations of sparse connections for fraud data. To address these three shortcomings, this paper proposes a fraud detection method, called NFE-GNN (Noise Filtering and feature Enhancement based Graph Neural Network method for fraud detection), to improve the fraud detection performance. The proposed NFE-GNN method first employs a dataset-based fraud rate sampling technology to achieve a balance of benign and fraudulent samples. Based on this, a parameterized distance function is introduced to calculate the similarities between nodes, and the optimal noise filtering threshold is obtained through adaptive reinforcement learning. Finally, an effective algorithm is presented to increase the connections between fraudulent samples, and enrich the topology information in the graph to enhance the feature representation capability of fraudulent samples. The experimental results on two publicly available datasets demonstrate that the detection performance of the proposed NFE-GNN method is better than that of state-of-the-art graph neural network methods.  
      Keywords:fraud detection;class imbalance;node classification;graph data;graph neural network;performance evaluation  
      52
      |
      58
      |
      3
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Federated Learning Scheme for Environmental Monitoring Based on Edge Computing

      JIANG Wei-jin, HAN Yu-qing, WU Yu-ting, ZHOU Wei, CHEN Yi-lin, WANG Hai-juan
      Vol. 51, Issue 11, Pages: 3061-3069(2023) DOI: 10.12263/DZXB.20230504
      Abstract:Aiming at the problems of unbalanced edge device resources, communication delay and low model quality in the field of environmental monitoring, this paper proposes an adaptive federated learning algorithm for environmental monitoring based on edge computing. This algorithm aims to use edge devices for data processing, and according to each the resource limitation of the device adjusts the aggregation frequency of the global model to better adapt to different monitoring environments. By considering the resource differences between edge devices, the algorithm adopts a strategy of dynamically optimizing the iteration frequency to improve the training effect of the model. Compared with the traditional fixed iteration frequency, the adjustment strategy of this algorithm is more flexible and can better adapt to different data distribution and participant characteristics. Through a large number of experimental evaluations, and using the same algorithm convolutional neural networks-federated learning (CNN-FL), federated averaging (FedAvg) and hierarchical federated edge learning (HFEL), the algorithm proposed in this paper has significant advantages in algorithm performance and economic cost. This algorithm provides an efficient, safe and reliable method for environmental monitoring. Expanded approach to data analysis and modeling to help drive improvements in environmental monitoring capabilities.  
      Keywords:environmental monitoring;adaptive federated learning;edge computing;model polymerization;optimization algorithm  
      106
      |
      88
      |
      1
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Resource Allocation Algorithm of AirComp Network Based on Multiple UAVs

      TAN Ling, XU Hai, LIU Yu-feng, XIA Jing-ming
      Vol. 51, Issue 11, Pages: 3070-3078(2023) DOI: 10.12263/DZXB.20230513
      Abstract:Over-the-air computation (AirComp) is an effective method to improve the efficiency of distributed data aggregation, which can complete some task calculations while transmitting in the air. Most existing researches focus on the single unmanned aerial vehicle (UAV) scheme, without considering the quality of data aggregation and the stability of the system, making it unsuitable for practical AirComp environments. Therefore, this paper proposes an AirComp network based on multiple UAVs collaboration, which aims to achieve the efficient data aggregation for multiple ground mobile sensors (GMSs). In order to refine data acquisition and fully reflect system status, a multi-constraint non-convex optimization problem is constructed to jointly optimize UAV-GMS association, the three dimensional (3D) deployment of UAVs, UAV denoising factors, and transmission power allocation, aiming for maximizing the system's minimum achievable rate. Giving the nonlinear characteristics of multiple constraints optimization problems, a deep deterministic policy gradient-based optimization algorithm for multiple UAVs cooperation in AirComp network (AirDDPG-UAV) is proposed to assist UAVs rapidly responding to aggregation missions in complex environments. A deterministic policy in deep reinforcement is adopted to optimize the states, behaviors, and rewards of the AirComp network, aiming to maximize the minimal achievable rate. The numerical results show that the AirDDPG-UAV algorithm can significantly improve the system's minimum achievable rate by more than 15% compared to the benchmark methods, while ensuring suitable system energy consumption and computational complexity. The AirDDPG-UAV algorithm also obtains satisfactory results in optimizing the mean MSE, which illustrates our method has excellent performance in scaling signals and thus is helpful for fast data aggregation. The experiments indicate the proposed scheme is appropriate for the distributed data aggregation with low cost and can obviously improve the efficiency and stability of data aggregation.  
      Keywords:unmanned aerial vehicle;over-the-air computation;3D deployment;deep deterministic policy gradient algorithm;ground mobile sensors;data aggregation  
      87
      |
      73
      |
      2
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Adaptive Spatial Sparsification for Efficient Multi-View Stereo Matching

      ZHOU Xiao-qing, WANG Xiang, ZHENG Jin, BAI Xiao
      Vol. 51, Issue 11, Pages: 3079-3091(2023) DOI: 10.12263/DZXB.20230353
      Abstract:To reduce the high computational complexity in constructing and aggregating cost volumes for multi-view stereo matching, existing methods commonly employ cascaded architectures or iterative optimization. However, these approaches still face two main challenges. The cascaded architectures narrow down the depth sampling range during the refinement stage, which may lead to erroneous estimation of depth discontinuities. While the inference time of iterative optimization networks linearly increases with the number of iterations, making it difficult to meet the requirements of real-time systems. To address these challenges, this paper proposes an efficient multi-view stereo matching network via adaptive spatial sparsification. We introduce a sparse matching cost volume that sparsely samples within the complete depth range, reducing computational complexity while maintaining the network's ability to model depth-discontinuous regions. Meanwhile, we propose a sparse iterative optimization method that progressively prunes regions with converged depth values during iterations using adaptive variational Dropout, resulting in sub-linear growth in inference time with iteration count. Experimental results on the public datasets, DTU and Tanks & Temples, demonstrate that the proposed method achieves 1.2× and 0.35× improvements of inference speed compared to CasMVSNet and PatchmatchNet, respectively. Moreover, it exhibits excellent performance in point cloud reconstruction, effectively handles details in depth-discontinuous regions, and demonstrates outstanding generalization capability.  
      Keywords:multi-view stereo;3D reconstruction;depth estimation;sparse neural networks;recurrent neural networks;Transformer  
      59
      |
      69
      |
      6
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Knowledge Graph Relation Path Network for Multi-Hop Intelligent Question Answering

      ZHANG Yuan-ming, JI Qi, XU Xue-song, CHENG Zhen-bo, XIAO Gang
      Vol. 51, Issue 11, Pages: 3092-3099(2023) DOI: 10.12263/DZXB.20230477
      Abstract:Complex multi-hop questions require knowledge reasoning to provide answers, which often involves integration of multiple pieces of knowledge to generate final answer. The existing knowledge graph (KG)-based multi-hop intelligent question answering methods often have complicated inference processes and do not consider structural and semantic information embedded in relation paths. To solve this problem, this paper proposes a knowledge graph relation path network for multi-hop intelligent question answering. It transforms the multi-hop intelligent question answering task into an optimization task of finding optimal relation path from KG. In this network, both the KG and question are embedded into low-dimensional vector spaces, and their unified vector representations are obtained. The topic entity and the question entity are combined to perform semantic scoring for generating candidate answers. Starting from the question entity and ending with candidate answers, a set of relation paths relevant to the question from the KG is extracted. The relation paths are further embedded into low-dimensional vector space to generate vector representations. By searching for the relation path with the highest semantic matching degree to the question in the vector space, the answer to the multi-hop question is generated. Experimental results on public datasets show that the proposed method has not only good performance but also good stability compared to the existing methods, and the performance does not decrease with the increase of problem hops.  
      Keywords:knowledge graph;intelligent question answering;complex multi-hop question;relation path;representation learning  
      130
      |
      137
      |
      5
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • A Multimodal 3D Object Detection Method Based on Double-Fusion Framework

      GE Tong-ao, LI Hui, GUO Ying, WANG Jun-yin, ZHOU Di
      Vol. 51, Issue 11, Pages: 3100-3110(2023) DOI: 10.12263/DZXB.20230414
      Abstract:The 3D object detection of camera and lidar multimodal fusion can comprehensively utilize the advantages of the two sensors to improve the accuracy and robustness of detection. However, due to the complexity of the environment and the inherent variability among multimodal data, 3D object detection still faces many challenges. In this paper, we propose a multimodal 3D object detection algorithm with a double-fusion framework. We design a voxel-level and grid-level double-fusion framework, effectively alleviating the semantic differences between modal data. We propose the ABFF (Adaptive Bird-eye-view Features Fusion) module to enhance the algorithm's ability to perceive small object features. Through voxel-level global fusion information to guide grid-level local fusion, we propose a Transformer-based multimodal grid feature encoder to extract richer context information in 3D detection scenes and improve the efficiency of the algorithm. The experimental results on the KITTI standard dataset show that the average detection accuracy of our proposed 3D object detection algorithm reaches 78.79%, which has better 3D object detection performance.  
      Keywords:deep learning;3D object detection;LiDAR;camera;multimodal information fusion  
      126
      |
      91
      |
      5
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • A Developer Recommendation Algorithm Based on Multi-Relationship Knowledge Enhancement

      DU Jun-wei, WANG Zhao-zhe, YU Xu, HU Qiang, JIANG Feng, GONG Dun-wei
      Vol. 51, Issue 11, Pages: 3111-3119(2023) DOI: 10.12263/DZXB.20230271
      Abstract:In recent years, crowdsourcing software development has gradually become an emerging software development model. However, with the continuous development of crowdsourcing platforms, the problem of information overload has become increasingly serious. It has become crucial to recommend suitable developers for tasks that are difficult to find reliable developers to complete on time. Traditional recommendation methods face two major challenges: First, the text features of tasks and developers are highly concise. Traditional recommendation methods focus on surface text information and fail to discover the large amount of knowledge entities contained therein. Second, tasks are one-time and this results in extremely sparse explicit interaction data. To solve these challenges, we propose a developer recommendation algorithm based on multi-relational knowledge enhancement. We identify entities and contextual entities from text features and link the relationship between tasks and developers from a knowledge perspective, uncovering the deep preferences of developers. In addition, we treat developers' participation in registering, submitting, and winning tasks as different preferences. We assign different weights to the relationships between tasks and developers and employ an attention mechanism to refine the importance of these different relationships. Finally, we enhance developers' feature representation using multi-relational neighborhood aggregation. We conduct extensive experiments on the real-world Topcoder dataset, and the results show that our method outperforms the best baseline method by an average of 11.7% in accuracy and 17.5% in mean reciprocal rank.  
      Keywords:developer recommendation;crowdsourced software development;multi-relationship;knowledge graph;graph neural network  
      61
      |
      89
      |
      0
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • A Large-Scale Constrained Multi-Objective Optimization Algorithm Based on Solution Space Reduction

      WANG Chao, HUANG Hui-tao, ZHANG Jing, QIU Jian-feng
      Vol. 51, Issue 11, Pages: 3120-3127(2023) DOI: 10.12263/DZXB.20230548
      Abstract:To tackle the challenges posed by high-dimensional and constrained solution spaces in large-scale constrained multi-objective optimization problems, this study employs an autoencoder-based solution space reduction technique to enhance the search efficiency of evolutionary algorithms. Firstly, a feasibility label pairing strategy is designed to train the autoencoder. By incorporating both feasible and infeasible solutions as two distinct classes of samples, a subspace can be constructed that captures the topological information of feasible regions. Also, this subspace can be regarded as the reduced representation of the original solution space. Secondly, the genetic operator is applied within the reduced subspace to produce the offspring, and the reconstructed outputs are subsequently mapped back to the original solution space using the decoder. This process can enable the location of the potential feasible regions. Lastly, an adaptive generation strategy is introduced to combine the advantages of offspring generated within both the reduced subspace and the original space, to prevent the model collapse and enhance the search efficiency. To validate the performance of the proposed algorithm, a comparative analysis is conducted against five state-of-the-art algorithms using publicly available test suites. The experimental results demonstrate that the proposed algorithm exhibits faster convergence speed and produces solutions of superior quality.  
      Keywords:large-scale constrained multi-objective optimization;evolutionary algorithms;auto-encoder;solution space reduction;offspring generation;feasibility  
      40
      |
      77
      |
      3
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Service Function Chains Scheduling of Information Networks Used for Emergency Treatment

      ZHANG Qing-hua, ZHANG Xian-chao, WANG Yin-hao, LU Jun
      Vol. 51, Issue 11, Pages: 3128-3136(2023) DOI: 10.12263/DZXB.20230293
      Abstract:To address the diverse emergency treatment needs, we investigate the scheduling problem of service function chains in healthcare networks. Firstly, emergency treatment services are categorized into four groups based on service area and urgency level, with each group assigned a corresponding priority and weight. We then propose a network framework, service model, and running process for healthcare networks. Secondly, we establish a service function chain scheduling model with the objective of minimizing the weighted total completion time. To address this problem at different scales, we develop a matching game algorithm and a Q reinforcement learning algorithm, both of which can yield an optimal scheduling scheme. Lastly, we conduct simulation experiments to assess our approach’s effectiveness, which confirms that our research can enable classified and graded emergency treatment services, and facilitate the development of intelligent medical and healthcare networks.  
      Keywords:healthcare networks;emergency treatment;service function chains scheduling;weighted total completion time;matching game;reinforcement learning  
      26
      |
      85
      |
      1
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Research on the Generation of Adversarial Samples Based on the Attribution of Principal Features

      WANG Shuo, XU Ru-zhi, GUAN Zhi-tao
      Vol. 51, Issue 11, Pages: 3137-3145(2023) DOI: 10.12263/DZXB.20230383
      Abstract:In order to improve the quality of adversarial training samples, this article explores the feature recognition process within deep learning models and proposes a transferability adversarial sample generation method based on main feature attribution. After extracting the main features of the samples, the algorithm attributes the features of the target layer neurons, and simplifies the gradient calculation by using the independence assumption. By inhibiting the recognition of active neurons, the algorithm can more efficiently obtain more transferable adversarial samples. Through a large number of experiments, compared to existing methods, the success rate of our algorithm in attacks against multiple models has been improved by at least 5%, it lays a foundation for further research on how to improve the robustness of the model.  
      Keywords:deep learning;image recognition;feature extraction;feature attribution;adversarial sample  
      15
      |
      70
      |
      2
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Task-Guided Radial Basis Function Network for Mitosis Detection in Breast Histopathology Images

      LIU Zi-yi, TANG Qi-ling, CAI Yu
      Vol. 51, Issue 11, Pages: 3146-3154(2023) DOI: 10.12263/DZXB.20230014
      Abstract:The number of mitosis is an important index to judge the malignancy of breast cancer, and it is of great significance in diagnosis, treatment and prognosis. However, in clinical practice, the detection of mitotic cells in breast cancer histopathology slices is mainly performed by manual counting by pathologists. This process is tedious, time-consuming and highly subjective. In this paper, we propose a two-stage method for automatic detection of mitosis in breast histopathology images. In the first stage, the localization of mitotic cell candidate patches is performed using a convolutional neural network as the backbone with deep supervision and attention mechanisms. In the second stage, the generated candidate patches are input into the verification model of radial basis function network to further improve the accuracy of mitotic cell identification. Aiming at the problem of large differences within mitotic cell types, this paper uses tasks to guide the definition of the radial basis function centers and uses the “local response” property of radial basis function to express the morphological diversity of mitotic cells. By evaluating on the weak label datasets ICPR 2014 and AMIDA 2013, the proposed network model has achieved the best F-score, which is 5.4% and 3.0% higher than the existing methods, thus proving that the validity of the method for mitosis detection.  
      Keywords:mitosis detection;radial basis function;K-means clustering;deep supervision;weak label  
      32
      |
      94
      |
      0
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • RGB-D Salient Object Detection Based on Cross-Modal Fusion and Boundary Deformable Convolution Guidance

      MENG Ling-bing, YUAN Meng-ya, SHI Xue-han, ZHANG Le, WU Jin-hua, CHENG Fei
      Vol. 51, Issue 11, Pages: 3155-3166(2023) DOI: 10.12263/DZXB.20230042
      Abstract:RGB-Depth (RGB-D) salient object detection is a meaningful and challenging task. The current method based on convolutional neural networks has achieved good detection performance in simple scenes, but cannot effectively handle scenes with cluttered background information, low-quality depth maps, and complex object contours. In order to solve the above problems, an RGB-D SOD model based on cross-modal fusion and boundary deformable convolution guidance is proposed in this paper. Firstly, the Swin Transformer is used as an extractor to extract features from the RGB modality and depth modality, respectively, which fuse the two modalities by using a cross-modal attention enhancement feature (CMAEF) module, to explore the common and complementary features of salient objects. Then, the proposed adjacent multi-scale feature enhancement (AMFE) module is embedded deep-level into the encoder to obtain rich global contextual feature information, which can locate the position of salient objects more accurately. Next, the boundary cue maps of salient objects are generated by boundary feature extraction decoder (U-Net architecture) constructed and repeated using cross-modal fusion features to ensure the integrity of the generated salient object boundaries. Finally, we designed a boundary deformable convolution guidance (BDCG) module that uses boundary cue maps with deformable convolution to guide the decoding of cross-modal fusion features to obtain more accurate saliency maps. Comprehensive experiments on six popular benchmark datasets compared with 25 mainstream methods demonstrate that the proposed model shows significant improvement in metrics, which proves the effectiveness of the proposed model.  
      Keywords:salient object detection;cross-modal fusion;boundary features;deformable convolution;saliency map  
      104
      |
      101
      |
      3
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • A Lightweight Music Recognition Method Based on Improved CRNN

      JIANG Ling-yun, JU Jin-heng, XU Jia, XIAO Fu
      Vol. 51, Issue 11, Pages: 3167-3175(2023) DOI: 10.12263/DZXB.20230031
      Abstract:The deep learning-based music score recognition method has improved recognition accuracy, while there is a dilemma of long single iteration time and multiple total iterations in model training. This work proposes CRNN-lite (lightweight Convolutional Recurrent Neural Networks) for music score recognition. CRNN-lite introduced residual depth separable convolution into the convolution layer, which reduced the computation and speeds up the feature map extraction. The bidirectional simple recurrent unit was used in the recurrent layer, and the strong dependence on serial computation was avoided by parallel computation. The parameters of the cross entropy function were adjusted at the transcription layer to learn unbalanced sample data. The results show that the proposed method improves the training speed, the single iteration time is 43% of the benchmark network, the symbol error rate is 1.12% and the sequence error rate is 14.5% on the distorted image data. The error rate indexes are better than the comparison scheme.  
      Keywords:optical music recognition;sequence recognition;convolutional recurrent neural networks;depthwise separable convolution;simple recurrent unit;unbalanced sample learning  
      92
      |
      99
      |
      3
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • JIANG Hong, MA Jiao-jiao, YAO Hong-ge, CHENG Si-yi, CHEN You, YU Jun
      Vol. 51, Issue 11, Pages: 3176-3186(2023) DOI: 10.12263/DZXB.20220617
      Abstract:When searching for a object, the human eye first roughly scans based on previous scanning experience to find potential locations for the object, and then conducts a detailed search. The former can be referred to as scanning based on temporal contextual information, while the latter can be referred to as searching based on location contextual information. Inspired by this, this paper proposes a rapid search method for small objects based on reinforcement learning that integrates spatio-temporal context information. The method builds a temporal context module based on a reinforcement learning search strategy to simulate the human eye's ability to obtain and utilize empirical information, then constructs an adaptive multi-scale window to extract location context information to simulate the human eye's ability to search carefully at possible locations. The two kinds of information cooperate alternately in the object search process to complete the object search. The experimental results show that the proposed algorithm brings around 2.9% gain on MS COCO benchmark, and can find an object within five search counts.  
      Keywords:small object detection;reinforcement learning;temporal context;location context;human eye saccade  
      37
      |
      105
      |
      0
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Chosen Ciphertext Combined Attack Based on Round-Reduced Fault Against SM2 Decryption Algorithm

      LI Hao-yuan, HAN Xu-cang, CAO Wei-qiong, WANG Jian, CHEN Hua
      Vol. 51, Issue 11, Pages: 3187-3198(2023) DOI: 10.12263/DZXB.20220481
      Abstract:SM2 algorithm is a commercial elliptic curve cryptographic algorithm designed by China. At present, the analysis of the implementation security of this algorithm usually follows the research results on the common components of elliptic curves rather than the structure and characteristics of the algorithm. At the same time, hash and verification steps in SM2 decryption algorithm make most of the fault attacks that need to exploit the error output not applicable. To solve this problem, according to characteristics of SM2 decryption algorithm, this paper proposes a chosen ciphertext combined attack that combines the round-reduced fault with side channel based on the idea of safe-error. The core of the attack is changing the number of rounds of scalar multiplication by fault injection, and determining the specific number of faulty rounds by side channel analysis. Then it constructs the chosen ciphertext based on partial key guesses combined with plaintext and correct ciphertext. And the chosen ciphertext is input to the decryption device with specific fault effect, verifying whether the partial key guess is correct by the output of the decryption device. Also, the applicability of the attack to different scalar multiplication methods and common protection countermeasures is analyzed in the paper. Lastly, we conduct practical attack experiments on the SM2 decryption algorithm with clock glitch injection and simple power analysis on an STM32F303 microcontroller chip based on the ARM Cortex M4. And we successfully recover the private key. The experimental results show that the attack method is feasible and practical.  
      Keywords:combined attack;round-reduced fault;side channel attack;chosen ciphertext;safe-error;SM2 decryption  
      106
      |
      72
      |
      1
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Semantic Segmentation of Remote Sensing Image Based on Multi-Scale Semantic Encoder-Decoder Network

      LIANG Yan, YI Chun-xia, WANG Guang-yu, HU Yue-hui
      Vol. 51, Issue 11, Pages: 3199-3214(2023) DOI: 10.12263/DZXB.20220503
      Abstract:This paper analyzes the existed processing scheme, and proposes a multi-scale semantic encoder-decoder networks (MSEDNet) by comprehensively using multiple technologies for the problems in remote sensing image semantic segmentation both multi-level information extraction and multi-scale feature diagram dependence characteristic. The MSEDNet consists of two parts: encoding part and decoding part. In the encoding part, the enhanced MobileNetV3 with residuals coordinate spatial attention (RCSA) is firstly proposed to extract semantic information, and then a multi-layer enhanced semantic context module (ESCM) is designed to improve representation ability of the multi-scale structure feature map. In the decoding part, a strengthen spatial detail information module (SSDIM) based on Multi-core Convolution and Focus Parallel is proposed to enhance the details and structural information of shallow features. Then triplet iterative multi-scale feature fusion (TIMSFF) strategy is designed to strengthen the multi-scale context fusion both deep global semantic information and shallow local detail features, for improving the segmentation accuracy. The proposed model has been experimentally verified on the ISPRS Vaihingen and Potsdam dataset. The overall segmentation accuracy (OA) reached 95.699% and 95.534% respectively, the mean F1-score (mF1) increased by 2.661% and 2.929% respectively, and the mean intersection over union (mIoU) increased by 3.973%and 4.012%, respectively. The number of param dropped to 6.77 M.  
      Keywords:remote sensing semantic segmentation;multi-scale semantic context;attention mechanism;spatial detail;multi-scale feature fusion  
      86
      |
      90
      |
      9
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Simulation Research on L Band Scatterometer System

      CHEN Zi-hao, LIN Wen-ming, MA Jian-ying, YUN Ri-sheng, ZHU Di, DONG Xiao-long, HE Yi-jun
      Vol. 51, Issue 11, Pages: 3215-3224(2023) DOI: 10.12263/DZXB.20220564
      Abstract:For the microwave remote sensing of sea surface salinity (SSS), the brightness temperature changes induced by the sea surface roughness are equivalent or even larger than those caused by the sea surface salinity. Consequently, the remote sensing of sea surface salinity should not only use L band microwave frequency, which is sensitive to SSS, for accurate brightness temperature measurement, but also use L-band scatterometer to estimate the sea surface roughness simultaneously. This plays a major role in improving the accuracy of sea surface salinity measurement. This paper studies the L band microwave scatterometer onboard the first Chinese ocean salinity mission. A simulation model is developed for the L-band multiple-beam push-brooming microwave scatterometer system. With the simulation model, the impact of radar pulse parameters, including the signal bandwidth and the pulse configuration of different polarizations on the system performance, is analyzed, and the design is optimized accordingly. The results show that under the baseline parameters of the L band scatterometer, the co-polarized backscattering measurement accuracy can be better than 0.1 dB when the pulse bandwidth is 1.5 MHz and the pulse repetition frequency is 100 Hz.  
      Keywords:L band;scatterometer;system simulation;measurement accuracy;communication error  
      81
      |
      74
      |
      0
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Special Video Recognition Based on Semantic Embedding Learning

      WU Xiao-yu, PU Yu-jiang, WANG Sheng-jin, LIU Zi-hao
      Vol. 51, Issue 11, Pages: 3225-3237(2023) DOI: 10.12263/DZXB.20220601
      Abstract:As special type of videos, violent video dissemination has become one of the hidden dangers facing the Internet environment, and intelligent recognition technology for violent videos is of great significance for maintaining Internet content security. Due to the diversity of collection sources, the distribution of violent videos usually shows large intra-class variance and small inter-class variance, and it is difficult for common violence recognition frameworks to adapt to complex and variable violent scenarios. Meanwhile, the word violence itself has highly abstract semantics, and it becomes a major difficulty to learn a generic semantic representation of violence from limited data. In response to these problems, we present a novel multimodal violent video recognition model based on semantic embedding learning. The model mainly consists of the following three parts. (1) Multimodal feature extraction. Considering that videos have multimodal properties, we use three different deep neural networks to extract feature representations of three modalities, i.e., appearance, motion, and audio. (2) Multimodal feature fusion. To obtain a robust universal video representation, a lightweight multimodal feature fusion module, referred to as MEFM (Multimodal Efficient Fusion Module), is designed in this paper. The module includes two parts: common space mapping and multimodal feature interaction, which can effectively suppress the interference between different modal information while fully interacting with multimodal features. (3) Semantic embedding learning. To accommodate violence datasets from different sources, we propose a multi-task learning method based on semantic embedding, which computes the semantic center of violence by introducing a center loss and uses cosine embedding loss to aggregate violent samples toward the center while discrete with non-violent samples to form a semantic discriminative feature representation, thus enhancing the generalization ability of the model and reducing the noise interference. Experiments on three publicly available datasets, VSD2015, Violent Flows, and RWF-2000, demonstrate that the violence video recognition framework proposed in this paper achieves competitive results by improving 4.79%, 0.81%, and 1.5% respectively, over the state of the arts.  
      Keywords:violent video recognition;multimodal feature fusion;semantic embedding;multi-task learning  
      45
      |
      95
      |
      0
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Anchor-Free Transformer Algorithm for Aerial Remote Sensing Target Detection

      YU Jiu-yang, HU Tian-hao, DAI Yao-nan, ZHANG De-an, XIA Wen-feng
      Vol. 51, Issue 11, Pages: 3238-3247(2023) DOI: 10.12263/DZXB.20220612
      Abstract:Aerial remote sensing image targets have the characteristics of multi-directional arrangement, small, and dense. The rotating target detection algorithm based on deep learning has the problem of poor detection accuracy. To solve this problem, the article proposes a novel anchor-free Transformer algorithm for aerial remote sensing target detection. Firstly, hierarchical Transformer is used to collect feature information of different resolutions to improve the range of feature information collection. Secondly, a new feedforward network (Spacial-FeedForward Neural network, SFFN) is constructed. SFFN combines the local space characteristics of 3×3 depth separable convolution with the global channel characteristics of multi-layer perceptron (MLP) to solve the shortcomings of feed forward neural network (FFN) in local space modeling. Finally, an anchor-free detector is built based on SFFN architecture, and the regression problem of prediction frame is divided into horizontal frame and rotating frame, which alleviates the loss discontinuity problem of rotating frame. The test results on DOTA dataset show that the average accuracy of this method has reached 75.83%, respectively, while achieving 92.47% of 5 small targets on NWPU VHR-10 dataset, which is more competitive in remote sensing target detection accuracy.  
      Keywords:remote sensing image;Target Detection;Transformer algorithm;anchor-free detector  
      71
      |
      66
      |
      6
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Removal of Impulse Noise in Grayscale Images Under Fuzzy Knowledge Measure

      GUO Kai-hong, CUI Ming-xi, LIU Ting-ting
      Vol. 51, Issue 11, Pages: 3248-3261(2023) DOI: 10.12263/DZXB.20220622
      Abstract:Aiming at the problems of inaccurate noise detection and unsatisfactory filtering effect in image impulse noise removal methods, this paper introduces a new theory of knowledge measurement, and proposes a new method of image denoising based on fuzzy knowledge. Firstly, the fuzzy knowledge measurement axiom system is improved and simplified, and a new fuzzy knowledge measurement parameterized model is proposed based on this. Secondly, the proposed model is used to calculate and analyze the similarity characteristics and the differences in the amount of knowledge between extreme points and neighboring pixels in the noise image, and effectively identify and distinguish impulse noise from ordinary extreme pixels. Finally, a maximum average correlation knowledge filter algorithm is proposed to achieve image denoising. The comparison experiments are carried out on the standard data set and the test image respectively, and the experimental results show that the proposed method can accurately identify the impulse noise, effectively improve the image filtering quality, and the main indicators and performance of the denoising results are significantly better than other similar algorithms. Compared with the average value of the comparison method, the index PSNR has an average increase of 7.8%, and the SSIM has an average increase of 10%. This paper applies the new theory of knowledge measurement to image denoising for the first time and achieves excellent results, which creates a new example for the innovative application of this theory in other related fields.  
      Keywords:knowledge measure;fuzzy set;image filtering;Impulse noise detection;adaptive window  
      35
      |
      65
      |
      1
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Electromagnetic Stealth Micro-Grid Metasurface Based on Particle Swarm Structure Optimization

      TU Jian-jun, MA Ding
      Vol. 51, Issue 11, Pages: 3262-3270(2023) DOI: 10.12263/DZXB.20220519
      Abstract:A new design method of metasurface's unit cell for getting preset reflection phase-frequency characteristics is proposed, in which the metal patch area of the cells are splitted into small grids, whose patches schemes are trained by particle swarm optimization (PSO) algorithm. This is an effective solution to resolve the problem in the design of unit cells' reflection phase of electromagnetic stealth metasurface, in which the metal patch area of a square unit cell is splitted into N×N grids (N is an even number, the free grids number n can be derived by N and full symmetric condition) , and the code string of the free grids' patch is trained by PSO, until the expected reflection phase-frequency characteristics (including the reflection phase's accuracy at target frequency and the flatness of phase-frequency curve in the neighborhood frequency band) obtain from the planar periodic structure constructed by the optimal unit cells. Taken as examples, in the centimeter band and terahertz band, four unit cells with a reflection phase difference of about 90° are given by PSO, and the 1-bit and 2-bit metasurfaces are constructed with phase gradients of 90° and 180° respectively, whose -10 dB RCS reduction effect near the target frequency are verified. It is verified that the unit cell with specific reflection phase can be designed automatically by this method, which is also universally applicable to different frequency bands, and the RCS reduction effect on a specific frequency and the bandwidth with satisfied reduction effect can both be obtained.  
      Keywords:electromagnetic stealth;metasurface;reflection phase;patch elements;particle swarm optimization  
      15
      |
      83
      |
      5
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • XIE Min, ZHAO Chuang, HU De-xiu, LIU Cheng-cheng
      Vol. 51, Issue 11, Pages: 3271-3281(2023) DOI: 10.12263/DZXB.20220799
      Abstract:To sort the pulse stream with a jagged pulse repetition interval (PRI) and reconstruct its PRI pattern, combined with the characteristics of jagger PRI pattern, a method that combines the first-order difference curve of difference of time of arrival (DTOA) and correlation analysis was proposed. Firstly, concerned DTOA terms were extracted based on the first-order difference curve of DTOA. Then, stagger PRI pattern was identified based on the autocorrelation and overlap rate of the concerned term, and the complete pattern was analyzed based on the cross-correlation of the concerned term. Finally, sequence retrieval was performed based on the analysis result. Experimental results show that the proposed method requires less computational effort and has excellent performance on sorting jagger PRI signal and reconstructing its PRI pattern, showing strong adaptability to pulse missing and great robustness to spurious pulse and the uniformity of the distribution of sub-periods.  
      Keywords:jagger PRI signal sorting;jagger PRI pattern recognition;jagged PRI pattern analysis;first-order difference curve of DTOA;correlation analysis  
      90
      |
      80
      |
      2
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • 3D Depth Point Cloud Supervision and Confidence Correction for Face Spoofing Detection

      HU Yong-jian, CAI Chu-xin, LIU Bei-bei, WANG Yu-fei, LIAO Guang-jun
      Vol. 51, Issue 11, Pages: 3282-3293(2023) DOI: 10.12263/DZXB.20220949
      Abstract:Due to feasibility and friendly user interaction, deep learning-based face recognition and identity authentication becomes one of the most popular artificial intelligence technologies in China. The face recognition and identity authentication system should secure that the captured face for verification is a living face rather than a fake face or called spoofing face. Otherwise, the output of the system is useless for business. Face spoofing detection or called living face detection mechanism is set in the front part of the system, and plays a key role in distinguishing a fake face from the input faces. Most current face anti-spoofing algorithms perform well in intra-dataset. However, the model training in lab is unable to simulate all aspects in the real-world application scenarios. As a result, the data distribution in source domain is not always similar to the data distribution in target domain, which causes the lab-trained algorithms barely perform as well as in lab. Although we can mitigate the performance degradation with the increase of detection feature types and dimensions, it tends to make the detection network very complex in structure and large in model size. In order to improve the generalization ability of model without resorting to large model, we design a face spoofing detection network using the 3D depth point cloud supervision and confidence correction scheme. The proposed approach consists of three major contributions. First, we design a shallow convolutional neural network called DenseBlockNet. It can well extract distinctive depth features between real faces and spoofing ones and has a small model size. Second, we establish the relationship between the 2D depth map produced by DenseBlockNet and the coordinates of sampling points, and thus create a 3D depth point cloud. We adopt the Chamfer loss to minimize the distance between the learned 3D depth point cloud and the ground truth 3D depth point cloud label, and use the binary cross entropy loss to supervise the difference between the learned 2D depth map and the ground truth 2D depth map label. Third, we introduce a prediction confidence map to correct the error of the learned 3D depth point cloud, so that it can avoid overfitting in intra-datasets and obtain good generalization ability in inter-datasets. Extensive experiments are conducted on 5 popular presentation attack databases, namely Reply-attack, CASIA-FASD, MSU-MFSD, Rose-Youtu, and OULU-NPU. Compared with 8 representative methods including 2 SOTA methods, the proposed method can achieve the least or second least half-total-error-rates in either intra-dataset or inter-dataset tests. Besides, it has the smallest model, the least amount of model parameters and the lowest computational complexity.  
      Keywords:face spoofing detection;3D depth point cloud;3D point cloud supervision;confidence correction;deep learning;generalization ability  
      55
      |
      83
      |
      0
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • ZHANG Xin-wen, LI Yue-heng, JIANG Yu-wei, JU Mei-yan, HUANG Ping
      Vol. 51, Issue 11, Pages: 3294-3304(2023) DOI: 10.12263/DZXB.20221160
      Abstract:In order to overcome the shortcoming of the traditional underwater weak turbulence model that cannot accurately fit the measured data, the generalized gamma distribution (GGD) which has been verified by series of laboratory experiments is chosen to characterize the weak oceanic turbulence. Based on this, a new aggregated fading channel model is proposed, which takes into account the GGD weak turbulence, zero boresight pointing error, and the implicit path loss and multipath propagation effect characterized by fading free impulse response (FFIR). Next, a closed-form expression of the hybrid fading probability density function integrating the GGD weak turbulence and zero boresight pointing error is derived using the higher transcendental Meijer-G function, and then, the theoretical closed-form expressions for the outage probabilities of the parallel and serial relaying underwater wireless optical communication (UWOC) systems are derived accordingly. Finally, Monte Carlo simulations are provided to verify the correctness of the above-derived analytical formulas for the outage probability and to investigate the effect of different system core parameters on outage performance of the systems. The results show that the introduction of relay nodes can effectively reduce the outage risk of UWOC over long distances. In contrast, the number of relay nodes, receiver aperture size, incident optical beam jitter variance, and other parameters can affect the outage performance. In addition, the FFIR delay spread caused by increasing of the divergence angle of the light source will lead to serious inter-symbol interference, which will result in a general decrease of 1~3 orders of magnitude in the outage performance of the relaying UWOC systems.  
      Keywords:generalized gamma distribution;weak oceanic turbulence;pointing error;relay;outage probability;underwater wireless optical communication  
      51
      |
      82
      |
      1
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • PU Xiu-juan, LIU Hao-wei, HAN Liang, REN Qing, LUO Tong-jun
      Vol. 51, Issue 11, Pages: 3305-3319(2023) DOI: 10.12263/DZXB.20221058
      Abstract:Alzheimer's Disease (AD) is a neurodegenerative disease that is currently incurable. Its accurate classification is advantageous to timely treatment and intervention at the early stage of AD, so as to reduce the incidence rate of AD and delay its progress. In this paper, one novel AD classification method utilizing deep learning and heterogeneous feature fusion is proposed. For the hippocampal structure in the brain, the three-dimensional lightweight multi-branch attention network (3D-LMBAN) is firstly constructed to extract the hippocampal depth features. Next, the three-dimensional multi-scale texture feature extraction method combining dual-tree complex wavelet transform (DTCWT) and gray-level run-length matrix (GLRLM) is proposed to extract hippocampal texture features. Then, the hippocampal volume and shape features are extracted by conventional methods. Finally, the dimension-reduction representation, concatenation and fusion of extracted various hippocampal features are performed using the constructed heterogeneous feature fusion network, and then AD classification is realized. The proposed AD classification method is evaluated on the EADC-ADNI dataset. The accuracy (ACC), F1 score and AUC of proposed AD classification method are 93.39%, 93.10% and 93.21%, respectively. The experimental results show that the proposed AD classification method is effective and better than other conventional AD classification methods.  
      Keywords:Alzheimer's Disease (AD);deep learning;attention mechanism;texture feature;feature fusion;hippocampus  
      50
      |
      62
      |
      2
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • YANG Xin, YANG Chun-ling
      Vol. 51, Issue 11, Pages: 3320-3330(2023) DOI: 10.12263/DZXB.20220958
      Abstract:Due to the lack of guidance from the parallel update theoretical solver framework, existing deep learning-based distributed compressed video sensing (DCVS) algorithms alternately use measurement values and reference frames to optimize the reconstructed non-key frame, resulting in the inability to fully combine the two types of information and limiting the quality of reconstruction. In order to solve this problem, this paper firstly uses Bayesian theory and maximum a posteriori estimation (MAP) to derive the optimization equation of non-key frame reconstruction in DCVS, and then derives the solution framework of the optimization equation based on the proximal gradient algorithm, including multi-information flow gradient update and aggregation equation. Based on it, this paper designs a multi-information flow gradient update and aggregation neural network module (MIGA), and constructs a deep multi-information flow gradient update and aggregation network (DMIGAN) for DCVS non-key frame reconstruction. MIGA uses the measurement values and multiple reference frames to update the current non-key frame by parallel gradients, and then performs information interaction and fusion, so as to fully combine multiple information flows to update and reconstruct the frame. In this paper, the MIGA and the denoising sub-network are cascaded to simulate a single iteration of the proximal gradient algorithm as the basic phase. The deep reconstruction network DMIGAN is constructed by cascading multiple phases to realize the deep optimization process of frame reconstruction. Experiments show that, compared with the representative traditional iterative optimization algorithm structural similarity based inter-frame group sparse representation (SSIM-InterF-GSR), the performance of DMIGAN is improved by 8.8 dB and 7.36 dB at low sampling rate and high sampling rate respectively; and compared with the representative deep learning reconstruction algorithm VCSNet-2, the performance is improved by 7.09 dB and 8.78 dB at low sampling rate and high sampling rate, respectively.  
      Keywords:distributed compressed video sensing;maximum a posteriori estimation;multi-information flow gradient update;information aggregation;iterative optimization;optical flow estimation  
      14
      |
      67
      |
      3
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Specific Emitter Identification Under Digital Pre-Distortion

      ZHAO Ya-qin, XIE Dan, WU Long-wen, DING Qin-yu, HAN Yi-shen, ZHANG Zheng-hua
      Vol. 51, Issue 11, Pages: 3331-3342(2023) DOI: 10.12263/DZXB.20221122
      Abstract:With the development of communication radar technology, new techniques such as pre-distortion have emerged to address the non-linear effects of radiation source transmitters, which weaken the individual characteristics of different radiation sources and thus deteriorate the individual source identification performance. To address the problem of reduced individual source identification under pre-distortion, this paper proposes an individual source identification model based on the SincNet filter structure. This paper uses the Grad-CAM method to analyse the residual network-like activation region and extract the co-occurrence matrix features for radiation source identification to verify the effectiveness of the local features of the signal after pre-distortion. This paper then proposes a SincNet filter structure-based algorithm for individual source identification, which reduces the computational effort while providing higher identification accuracy at low signal-to-noise ratios. The negative effect of digital pre-distortion on the individual identification of radiation sources is verified experimentally and the results on the measured data show that the individual identification rate of the proposed method reaches 94% at a signal-to-noise ratio of 0 dB, which is a significant improvement compared to other advanced individual identification algorithms in this paper.  
      Keywords:specific emitter identification;pre-distortion;Grad-CAM;SincNet  
      74
      |
      78
      |
      1
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Blind Signal Separation Combining Impulse Response Remodeling and Expectation Maximization

      XIE Yuan, ZHANG Xu, ZOU Tao, MA Ge, YU Jin-shi, SUN Wei-jun
      Vol. 51, Issue 11, Pages: 3343-3353(2023) DOI: 10.12263/DZXB.20230272
      Abstract:The separation of multichannel underdetermined convolutive speech mixing signals is a difficult problem in the field of blind signal separation. Due to the acoustic echo and reverberation in the mixing signals, it is difficult to completely and clearly separate the real source signals. Most traditional blind signal separation algorithms are suitable for low reverberation, while in high reverberation scenarios, the separation performance of the algorithm rapidly degrades or even fails. To separate multichannel convolutive mixing signals with acoustic echo and reverberation, a blind signal separation algorithm is proposed combining impulse response remodeling and expectation maximization, which exhibits good separation performance in both low and high reverberation environments. Firstly, a pre-filter is designed using impulse response remodeling technology based on infinite and p-norm to eliminate audible echoes, remodeling the room impulse response and improving the quality of the mixing signals. Then, a hierarchical clustering method is used to estimate the mixing matrix for the remodeled mixing signals, the new model parameters real-time update rules are designed based on expectation maximization algorithm framework, and the source signals are reconstructed combining the impulse response remodeling and expectation maximization. Experimental results show that the proposed algorithm can effectively separate the speech mixing signals with acoustic echoes in different reverberation environments, owning superior separation performance and good robustness to noise.  
      Keywords:blind signal separation;impulse response remodeling;multichannel convolutive mixtures;expectation maximization  
      25
      |
      110
      |
      0
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Parallel Multi Pipeline Design of Embedded Heterogeneous AI Computing Systems

      ZHAO Er-hu, WU Ji-wen, XIAO Si-ying, JIN Zhen-jie, XU Yong-jun
      Vol. 51, Issue 11, Pages: 3354-3364(2023) DOI: 10.12263/DZXB.20230281
      Abstract:Due to the limited power consumption and the need for real-time intelligent processing of multiple sensors, embedded AI computing systems desire for higher energy efficiency and more parallel intelligent computing services simultaneously.The digital signal processing architecture DSP+FPGA commonly used in traditional embedded computing systems is not suitable for multiple ANN models inference acceleration.Based on embedded heterogeneous intelligent computing architecture ARM+DLP+SRIO, this paper proposes a parallel multi pipeline design method by taking advantage of the characteristics of multi chip, multi-core and multi memory channels of deep learning processors.Considering the time cost of data transmission, copy, reference and feedback, this method allocates intelligent computing resources for different neural network models to achieve the maximum end-to-end throughput.The experimental results show that the utilization of the deep learning processor using the parallel multi pipeline design method is about 25.2% higher than that of a single pipeline, and about 30.7% higher than that without pipeline.It meets the real-time intelligent processing requirements of visible light, infrared and SAR images, and is valuable for practical applications.  
      Keywords:embedded AI computing systems;heterogeneous computing architecture;neural network model;parallel multi pipeline;deep learning processor (DLP)  
      92
      |
      83
      |
      1
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • A Semantic Segmentation Method of Embryo Image Based on Curriculum Learning

      TANG Hong-zhong, WANG Wei, WANG Tao, LU Wang-da, HUANG Xiang-hong, ZHANG Jin
      Vol. 51, Issue 11, Pages: 3365-3376(2023) DOI: 10.12263/DZXB.20230097
      Abstract:Morphological features of embryo play an important role to evaluate the quality of human embryos in vitro fertilization. Embryologists mainly use time-lapse imaging (TLI) technology to observe the morphological variation of embryo images and select the most potential embryos for subsequent transfer or cryopreservation. However, manual evaluation is not only time-consuming and laborious, but it also demands specialty-oriented skills and has some subjectivities. To solve this problem, we propose a semantic segmentation method of embryo images based on curriculum learning to segment embryo cell, cytoplasm and pronucleus, which can provide the quantitative parameters of morphological feature for the subsequent evaluation of embryo quality. First, IoU (Intersection over Union) index that is usually used to evaluate the performance of semantic segmentation algorithms is adopted to construct a difficulty scoring function (SF) of curriculum learning. All samples are sorted from easy to difficult according to SF value. Second, we define a pace function (PF) by combining the SF and the number of target categories, and sample subsets is established with increasing sequence of difficulty. Lastly, we design a multi-stage progressive U-net (MSPU) model to segment embryo cell, cytoplasm and pronucleus embryo images at pronuclear-stage, in which the network at different stages are trained using the sample subsets with increasing sequence of difficulty. Experimental results demonstrate that our proposed MSPU model obtains a satisfactory performance on the semantic segmentation of embryo images, and IoU is improved by 1.4% compared with Vanilla Baseline. Our proposed MSPU model shows a pleasing consistency on the easy and difficult segmentation tasks, for example the improvement of 4.6% and 1.2% can be obtained for the segmentation of embryonic cells and pronucleus, respectively.  
      Keywords:curriculum learning;embryo time-lapse image;multi-stage progressive U-net;semantic segmentation;scoring function;pacing function  
      42
      |
      69
      |
      1
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • MoGE: Graph Context Enhanced Multi-Task Recommendation Method

      ZHOU Jun, HU Bin-bin, ZHANG Zhi-qiang, CHEN Chao-chao
      Vol. 51, Issue 11, Pages: 3377-3387(2023) DOI: 10.12263/DZXB.20220964
      Abstract:Multi-task learning (MTL), which jointly tackles multiple tasks through information sharing, has been extensively used in a variety of recommendation applications. Recently, current efforts targeted for recommendation focus on learning task relationships based on the multi-gate mixture-of-experts (MMoE) architecture with shared input features (i.e., subtle feature engineering for user-item interaction). Recent evidences suggest the graph neural network (GNN) as a powerful component in characterizing deep interaction context for recommendation, greatly contributing to easing the data sparseness issue in online advertising services. Hence, we make the first attempt to explore the GNN towards multi-task recommendation, by designing mixture of graph enhanced expert networks (MoGE). Specifically, we propose a novel multi-channel graph neural network to jointly model high-order information with the user-item bipartite graph as well as derived collaborative similarity graphs for users and items. A group of graph augmented expert networks are introduced on top of the learnt deep interaction context for cooperatively contributing to the multi-task recommendation. Experimental results on three real-world datasets show that MoGE consistently and significantly outperforms state-of-the-art baselines across all target tasks.  
      Keywords:multi-task learning;graph learning;recommender system;attention mechanism;deep learning  
      43
      |
      62
      |
      4
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • Research on the Online Detection for the Entropy Source Based on the Cycle Count Difference

      OU Qing-yu, LUO Fang, CHU Wei-yu
      Vol. 51, Issue 11, Pages: 3388-3396(2023) DOI: 10.12263/DZXB.20220119
      Abstract:As the key element of the cryptography application, the true random number plays an irreplaceable role. To guarantee its quality, it can be mainly generated by the entropy source, composed of the random physical processes, so it is vulnerable to the ambient interference and the attack, and thus the security of the cryptography application can be threatened. To guarantee the quality of the true random number, the current main international standards require the true random number generator (TRNG) to be provided with the online inspection on the generated raw random numbers. However, the current online inspection is mainly implemented by the method of sampling statistics, so it has the problems of complicated implementation and huge resource consumption. On the other hand, because of the statistic model bias and the disturbance on the entropy source output distribution, the excessive estimation has become a common problem in the current online inspection method. In this paper, the current widely applied ring oscillator-based entropy source is researched. By analyzing the cause and the impact of the entropy source characteristic variation, the heterogeneous isomorphism cycle count difference is prompted to characterize the entropy source, and with the calibration of the variant threshold value, the online inspection method, with high accuracy and strong real-time, is established. Compared with the current online inspection methods, the proposed method has significant advantages of resource consumption, and the variation of the noise components, as well as the disturbance inductive couplings, can be reflected in real time, so the inspection on the characteristics of TRNG, can be effectively implemented.  
      Keywords:true random number generator;entropy source;online detection;disturbance in environment;cycle count;Allan variance  
      29
      |
      77
      |
      0
      <HTML>
      <L-PDF> <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08
    • 中英文目录

      Vol. 51, Issue 11, Pages: 3397(2023)
      7
      |
      30
      |
      0
      <HTML>
      <Meta-XML>
      <Citation><Bulk Citation>
      Updated:2025-12-08