Journal of Zhengzhou University(Natural Science Edition)

 

2025年04期目次

  • Multi-person Pose Estimation Algorithm in Complex Scenes

    SHI Lei;WANG Tianbao;MENG Caixia;WANG Qingxian;GAO Yufei;WEI Lin;School of Cyber Science and Engineering, Zhengzhou University;School of Computer and Artificial Intelligence, Zhengzhou University;Department of Image and Network Investigation, Zhengzhou Police University;

    The cross-obscuration of individuals in complex scenes led to low accuracy and incorrect skeleton connections in existing human pose estimation algorithms. Therefore, a multi-person pose estimation optimization algorithm in complex scenes was proposed. Firstly, the ordinary convolution was replaced with the grouped cascade convolution, which was combined with feature fusion to promote the exchange of information between channels. The accuracy of the algorithm was improved without incurring additional computational costs. Secondly, the spatial attention mechanism was introduced to mine the spatial semantic features related to the human pose estimation task, and the network structure was parallelized to enhance the performance of the algorithm. Finally, the embedding positions of the large convolutional kernel and the attention mechanism were lightweighted to reduce temporal overhead. Compared to the existing bottom-up pose estimation algorithm OpenPifPaf++, the proposed algorithm improved the average accuracy by 0.8 percentage points on the COCO 2017 dataset. Compared with the OpenPifPaf algorithm, the proposed algorithm improved the average accuracy by 1.2 percentage points on the CrowdPose dataset, and the corresponding accuracy for complex scenes by 1.5 percentage points.

    2025 04 v.57 [Abstract][OnlineView][Download 496K]

  • UAV Game Path Planning Based on Deep Reinforcement Learning

    XUE Junxiao;ZHANG Shiwen;LU Yafei;YAN Xiaoran;FU Wei;School of Cyber Science and Engineering, Zhengzhou University;Institute of Artificial Intelligence,Zhejiang Lab;Research Center of Space Computing,Zhejiang Lab;

    A deep reinforcement learning model driven by knowledge and data was proposed to address the low learning efficiency of deep reinforcement learning methods in complex environments for unmanned aerial vehicle(UAV) game tasks. Firstly, drawing on the idea of imitation learning, a genetic algorithm was employed as a heuristic search strategy, and expert experience knowledge was collected. Secondly, the UAV interacted with the environment through deep reinforcement learning and collected online experience data. Finally, a deep reinforcement learning model driven by knowledge and data was constructed to optimize UAV game strategies. Experimental results indicated that the proposed model effectively improved the convergence speed and learning stability, and the trained agents demonstrated better autonomous game path planning capabilities.

    2025 04 v.57 [Abstract][OnlineView][Download 417K]

  • An eBPF-empowered Task Offloading Approach for Cloud-edge-end Collaborative Networks

    LI Shuo;YAN Fei;ZHANG Liqiang;LUO Qingcai;YANG Xiaolin;School of Cyber Science and Engineering, Wuhan University, Key Laboratory of Aerospace Information Security and Trusted Computing of Ministry of Education;Shandong Inspur Science Research Institute Co., Ltd;Inspur Smart Technology Co., Ltd;

    As a key enabling technology for cloud-edge-end collaborative networks, computing offloading is an effective approach to alleviate issues like insufficient computing capabilities and limited resources in edge embedded devices. Some existing studies focused primarily on reducing latency and energy consumption in simulated settings. Yet accurately perceiving the real-time dynamics of cloud-edge-end collaborative networks and implementing flexible task offloading strategies remained an urgent challenge to tackle. FreeOffload, a task offloading framework for Cloud-Edge-End Collaborative Networks was proposed. Leveraging eBPF technology, FreeOffload realized real-time awareness of computing resources and network status. It also incorporated flexible task re-offloading schemes tailored for heterogeneous embedded end devices, which achieved load balancing across edge nodes. A small-scale cloud-edge-end prototype tested for evaluation was constructed. Results demonstrated that FreeOffload while efficiently and flexibly offloaded tasks from end devices, with low overhead.

    2025 04 v.57 [Abstract][OnlineView][Download 849K]

  • A Privacy-preserving Federated Learning Framework Based on Consortium Chain

    WEI Chao;YANG Wenshao;LIU Wei;School of Science, Yanshan University;School of Cyber Science and Engineering, Zhengzhou University;

    Aiming at the shortcomings of existing federated learning models in privacy protection and poisoning attack defense, a privacy-preserving federated learning framework based on consortium chain was proposed. Firstly, the framework employed homomorphic encryption techniques and Laplacian noise to ensure data privacy, effectively preserving the confidentiality of data from various parties during model training. Secondly, through the consensus protocol of the consortium chain and a model aggregation algorithm, distinct gradient aggregation weights were assigned to different participants, mitigating the impact of malicious parties on model aggregation and enhancing the robustness of the model. The experimental results conducted on the MNIST and Fashion-MNIST datasets demonstrated that even with a malicious participant ratio up to 40%, the proposed framework could still maintain high model accuracy with label reversal attack and backdoor attack.

    2025 04 v.57 [Abstract][OnlineView][Download 311K]

  • Anchor Multi-view Clustering Based on Adaptive Fusion of Global and Local Information

    RAN Zhuang;WANG Siwei;ZHU En;School of Computer, National University of Defense Technology;

    Subspace-based multi-view clustering algorithms have attracted much attention due to their good clustering performance and mathematical interpretability. Among them, some large-scale multi-view subspace clustering algorithms based on anchor strategy can effectively reduce the spatiotemporal complexity. However, existing algorithms often learned the subspace self-representation matrix from the global structure, ignoring the local structure information between the view data, anchors and the subspace self-representation matrices. Inspired by the multi-view self-weighted multi-graph learning algorithm, the anchor multi-view clustering based on adaptive fusion of global and local information(AMVC-AFGL) algorithm was proposed. The proposed algorithm aimed to learn a more effective subspace anchor graph matrix for each view data by adaptively allocating view weights and fusing the global information and local information between the data, and then concatenated them into a smaller fusion anchor graph matrix for spectral clustering. Extensive experiments were carried out on 10 public real benchmark datasets, and compared with other 12 advanced multi-view clustering algorithms, the results showed the effectiveness and scalability of the proposed algorithm.

    2025 04 v.57 [Abstract][OnlineView][Download 417K]

  • The Traffic Flow Prediction Algorithm Based on Variational Autoencoder

    CUI Wenyuan;TENG Fei;HE Baisheng;HU Xiaopeng;QIU Ge;School of Computing and Artificial Intelligence, Southwest Jiaotong University;Sichuan Port Investment Southern Sichuan Port Industry Investment (Group) Co., Ltd.;

    In order to solve the problem that the existing traffic flow prediction models could not fully mine the spatio-temporal dependence of complex and dynamic traffic flow data, a traffic flow prediction model based on variational autoencoder(AST-VAE) was proposed. Firstly, the variational inference and residual decomposition mechanism were used to separate the traffic flow signal into hidden diffusion signal, intrinsic signal and random signal. The temporal and spatial correlations in the three signals were then extracted using different learning modules. Finally, the three multi-dimensional features were fused to capture the global spatio-temporal dependence. With two real traffic datasets, the effectiveness and feasibility of the specific modules of the model were analyzed, and the experimental results showed that AST-VAE was always better than the existing models in the traffic flow prediction task, and the error was low, and it had good prediction performance.

    2025 04 v.57 [Abstract][OnlineView][Download 897K]

  • Infrared Image Enhancement Algorithm Based on Edge

    CHEN Ming;MA Guoqiang;HUANG Wanwei;GAO Tieliang;LI Yuhua;NIU Yanfei;College of Software Engineering, Zhengzhou University of Light Industry;Bussiness School, Xinxiang University;

    In recent years, the low latency and high efficiency characteristics of edge computing have extensive applications in infrared imaging systems, which could effectively reduce operational costs. However, issues such as low contrast and blurry details in infrared images still needed to be addressed. To solve these problems, an edge infrared image enhancement algorithm based on Lagrange interpolation and multi-scale guided filtering was proposed. This algorithm consisted of two phases. In the first phase, the Lagrange interpolation algorithm was used to achieve non-uniform correction for infrared data in the edge end. The Lagrange nonlinear interpolation was more in line with the response curve of the infrared detector, which effectively solved the problem of non-uniform noise introduced during imaging. In the second phase, multi-scale guided filtering was employed to process the infrared image in a hierarchical manner. Multiple scales were used to extract various details from images, and by fusing these different detail layers, a richer detailed information was obtained. Experimental results demonstrated that, compared to 5 traditional algorithms, this algorithm outperformed the suboptimal algorithm by 15.2% in the enhancement measure evaluation metric and achieved a 7.9% improvement in the peak signal to noise ratio metric.

    2025 04 v.57 [Abstract][OnlineView][Download 388K]

  • Regret Bounds for Online Pairwise Learning with Predictable Non-convex Loss Functions

    LANG Xuancong;WANG Mei;LIU Yong;LI Chunsheng;College of Computer and Information Technology, Northeastern Petroleum University;Heilongjiang Key Laboratory of Petroleum Big Data and Intelligent Analysis, Northeastern Petroleum University;Gaoling School of Artificial Intelligence, Renmin University of China;Beijing Key Laboratory of Big Data Management and Analysis Method at Renmin University of China;

    Online pairwise learning is a machine learning model in which the loss functions depend on a pair of instances. Generalization is an important aspect of online pairwise learning theory research. Most of the existing works on online pairwise learning used adversarial loss functions and provided regret bounds only with convex loss functions. However, convexity was not typically applicable in practical scenarios. For non-convex online pairwise learning, the regret bound of online pairwise learning with a "predictable" loss function based on stability analysis was provided and the corresponding stability analysis was conducted. Through the relationship between stability and regret, a common way to measure the generalization ability of online pairwise learning, the regret bound was established with a "predictable" non-convex loss function. It was proved that when the learner obtained an offline oracle, "predictable" non-convex generalized online pairwise learning reached the regret bound of O(T ~(-3/2)). This study enriched the theoretical research on non-convex online pairwise learning and was superior to the existing theoretical guarantees.

    2025 04 v.57 [Abstract][OnlineView][Download 235K]

  • Three-stage Guillotine Cutting Problem Based on Column Generation and Genetic Algorithm

    JI Ronggen;HU Zhihua;TIAN Xidan;WEI Yuehe;Logistics Research Center, Shanghai Maritime University;School of Economics and Management, Tongji University;

    In industrial customized production, the three-stage guillotine cutting layout method is often used, and a complete product must be cut after three stages. To address this issue, a mixed integer programming model for two-dimensional guillotine cutting of sheet metal was established with the goal of maximizing sheet metal utilization. The three-stage cutting problem was abstracted as a sorting problem with size constraints. In the first stage, two cutting methods were used: horizontal and vertical cutting. The cut product items could be placed at 0° or 90°, and the subsequent two stages of cutting must meet the requirement of no overlap between any two product items. To improve solution efficiency, the model was decomposed into a main problem and several sub problems, and an iterative re-optimization framework and algorithm for genetic algorithm and column generation were proposed to solve the problem. In each iteration, the genetic algorithm could provide multiple columns that met the conditions and added them as new columns to the main problem. In addition, the algorithm could re-optimize solutions with low utilization of sheet metal, improving the possibility of transforming the inferior solution to the optimal solution. The experimental results showed that in small-scale examples, the model could obtain accurate solutions. The proposed algorithm achieved a plate utilization rate of over 85% in large-scale examples, with a short solving time and could meet the requirements of industrial production.

    2025 04 v.57 [Abstract][OnlineView][Download 259K]

  • Minimum Conflict Heuristic Assisted Discrete Ocean Predator Solving RB Model

    YANG Yi;WANG Xiaofeng;HUA Yingying;YANG Lan;PANG Lichao;School of Computer Science & Engineering, North Minzu University;The Key Laboratory of Images & Graphics Intelligent Processing of State Ethnic Affairs Commission,North Minzu University;

    The revised B(RB) model was a stochastic instance model that possessed an exact phase-change growth domain in constraint-satisfiable problems. A solution algorithm was proposed for solving RB model instances, based on a combination of meta-heuristics and local search. Utilizing the marine predator algorithm, the initial solution space was discretized by real coding, and the three core phases of the marine predator algorithm were optimized. The current candidate solutions were targeted to guide the search towards the optimal solution. In the final stage, with the assistance of the local search method, the current optimal solution was passed to the minimum-conflict heuristic of the annealing strategy when the resulting optimal solution failed to satisfy the solution of the RB model instances, further enhancing the algorithm′s solving efficiency. Experimentally, the algorithm was shown to be significantly more accurate and time-efficient than many other heuristic algorithms. It demonstrated the potential of high probability solution even when it was close to the satisfiability phase transition point.

    2025 04 v.57 [Abstract][OnlineView][Download 566K]

  • Multi-case Derivational Adaptation Mechanism for Intelligent Knowledge Service

    ZHANG Jianhua;WEN Dandan;CAO Ziao;GAO Yarui;School of Management,Zhengzhou University;

    The proposed mechanism aimed to enhance autonomy and accuracy in adapting case knowledge within intelligent knowledge services. It utilized a multi-case induced adaptation approach that hinged on the correlation between decision attributes. Initially, the mechanism assessed the relevance of decision attributes using probability theory. Then, it employed different adaptation methods based on the independence or relevance of these attributes. When decision attributes were correlated, the mechanism selected adaptation techniques such as power set-based adaptation, adapter chain-based adaptation, and adapter chain combination-based adaptation. These choices depended on variations in the number of decision attributes and the size of the value domain. Simulation experiments conducted on the Student Performance-MAT dataset demonstrated the effectiveness of the mechanism.

    2025 04 v.57 [Abstract][OnlineView][Download 161K]

  • Optimal Location Selection Algorithm for Fusing Semantic Information

    LI Menghan;YAN Yantong;LI Lihong;Department of Science, North China University of Science and Technology;Hebei Province Key Laboratory of Data Science and Application;Tangshan Key Laboratory of Engineering Calculation;

    Currently, the widespread application of location services poses challenges to personal information security. In response, researchers explored various strategies for location privacy protection, with algorithms incorporating semantic information emerging as a key focus. A new idea for selecting location centers was proposed using semantic information, integrating this concept with semantic distance to choose the optimal anonymous candidate set, thereby ensuring the semantic and physical diversity of the location set. Experimental results demonstrated that compared with DLS, Enhanced-DLS, and MMDS algorithms, the method maintained diversity in both semantic and physical aspects, effectively reducing the accuracy of location data and protecting user privacy.

    2025 04 v.57 [Abstract][OnlineView][Download 277K]