Table of Content

10 September 2026, Volume 46 Issue 9

Previous Issue   

Artificial intelligence

Maximum entropy reinforcement learning method with temperature coefficient adaptive adjustment

Tao XU, Bin HU, Jin QIN

2026, 46(9):  2725-2731.  DOI: 10.11772/j.issn.1001-9081.2025081006

Asbtract ( )   HTML ( )   PDF (1760KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Maximum entropy reinforcement learning has attracted considerable attention due to its exceptional exploration capabilities in complex tasks. The temperature coefficient， which regulates the importance of the policy entropy term， is a key factor in balancing policy exploration and exploitation， and its setting affects algorithm performance significantly. However， the existing methods for adjusting temperature coefficient are typically based on empirical presets or fixed target entropy adjustments， neglecting state-related variations in exploration and lacking effective adaptive mechanisms. Therefore， a state-based adaptive temperature coefficient adjustment method was proposed. In the method， a neural network model was employed to predict an appropriate temperature coefficient based on a given state， and normalized temporal-difference error was used to construct supervisory information to guide model training， thereby realizing adaptive adjustment of entropy term weights based on the state. Then， this adaptive temperature coefficient adjustment method was integrated with the SAC （Soft Actor-Critic） algorithm to form a state-based temperature coefficient adaptive adjustment SAC algorithm. Experimental results show that on standard MuJoCo control tasks， this algorithm generally outperforms the baseline algorithms such as SAC and TD3 （Twin Delayed Deep Deterministic policy gradient） in terms of performance and training stability， during the evaluation stage， the algorithm has the average return increased by at least 4.3% and 6.2% compared with SAC and TD3 algorithms， respectively， validating the effectiveness of the state-based adaptive adjustment method for temperature coefficient.

CNN pruning and quantization compression method for reconfigurable structures

Yixin ZHANG, Lin JIANG, Yuancheng LI, Chen JI

2026, 46(9):  2732-2740.  DOI: 10.11772/j.issn.1001-9081.2025081055

Asbtract ( )   HTML ( )   PDF (1728KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the problems of high memory access overhead， redundant computation， and limited efficient deployment caused by the large parameter size of Convolutional Neural Networks （CNNs）， a CNN pruning and quantization compression method for reconfigurable structures was proposed to integrate network structure characteristics and hardware deployment requirements， and to perform co-optimization from both pruning and quantization dimensions. First， a convolutional layer pruning strategy based on feature similarity was introduced， in which feature evaluation， cluster and grouping， similarity calculation， and redundancy removal were performed in turn to filter out low-contribution and redundant filters. Second， progressive threshold pruning was applied at the fully connected layers to compress redundant weights. Third， in the quantization part， layer sensitivity indices were constructed using Hessian traces， and the precision of each layer was assigned adaptively under a bit-width budget. Finally， combining the characteristics of reconfigurable structures， an optimized deployment scheme was designed. Experimental results on the CIFAR-10 dataset show that the proposed method achieves a compression ratio of 16.2x for VGG16， surpassing Automated deep neural network Pruning and Quantization framework （APQ） （13.9x）. Compared to the model using fixed 16-bit precision， the pruned VGG16 using the proposed deployment scheme on a self-reconfigurable and self-evolvable Artificial Intelligence （AI） chip has the inference latency reduced from 23.3 ms to 9.1 ms， achieving a 2.56x speedup. It can be seen that the proposed method reduces the storage and transmission costs while maintaining the classification accuracy， improving the deployment efficiency and computational performance on edge devices.

Federated learning algorithm based on prototype clustering and Fisher information matrix weighted fusion

Xingyao WANG, Xuebin CHEN

2026, 46(9):  2741-2751.  DOI: 10.11772/j.issn.1001-9081.2025081017

Asbtract ( )   HTML ( )   PDF (851KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Federated learning， as an important research direction in the field of artificial intelligence， constructs global models through distributed collaborative training under the premise of ensuring data privacy， providing a new paradigm for joint modeling in privacy-sensitive scenarios. However， the existing methods are challenged by performance degradation， training oscillation， and slow convergence when faced with heterogeneous client data. Therefore， a federated learning algorithm based on prototype clustering and Fisher information matrix adaptive weighted fusion， named FedPFA， was proposed. First， the prototypes uploaded by clients were clustered by the server to mitigate data distribution discrepancies and enhance the consistency and generalization capability of the global model. Subsequently， differentiated weights were assigned to the prototypes by incorporating the trace estimated values of the Fisher information matrix from each client， thereby highlighting the contributions of high-quality clients and suppressing interference from noisy or inadequately trained clients during the fusion process. Experimental results on the MNIST， Fashion-MNIST， and CIFAR-10 datasets demonstrate that， under the setting of 20 clients and a participation rate of 0.3， FedPFA achieves accuracy improvements of 19.64， 26.14， and 16.15 percentage points， respectively， compared to the Federated Global prediction Header （FedGH） algorithm. It can be seen that FedPFA ensures performance enhancement， and improves the convergence speed and stability of the global model significantly， exhibiting strong robustness and practical application value on various datasets.

Heterogeneous enhancement and multi-source knowledge fusion for aspect-based sentiment analysis

Yifan WANG, Hu HAN, Dong LI, Yating FAN, Lin LI

2026, 46(9):  2752-2760.  DOI: 10.11772/j.issn.1001-9081.2025080979

Asbtract ( )   HTML ( )   PDF (940KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Aspect-Based Sentiment Analysis （ABSA） is an important task in natural language processing. The existing methods generally adopt single-type nodes to model semantic associations between aspects and opinion terms， failing to distinguish the contributions of different node types to sentiment representation effectively. Moreover， current models usually construct syntactic trees using dependency parsers for semantic analysis， ignoring fine-grained semantic information of the word itself and implicit sentiment intensity modification relations in the text. To address the above issues， a Heterogeneous enhancement and Multi-Source knowledge fusion Graph Convolutional Network （HMS-GCN） model was proposed. First， a heterogeneous graph with differentiated constraints was constructed to capture semantic and sentiment information in sentences multi-dimensionally by using the relations of multi-type nodes and their corresponding edges in the heterogeneous graph， and importance of different types of nodes and their internal components were focused on by utilizing a dual-channel attention mechanism. Second， sememe knowledge was adopted to enhance the semantic representation of initial textual features， and hierarchical phrase structural information was integrated on the basis of the syntactic adjacency matrix， so as to improve the model’s syntactic dependency comprehension. Finally， the heterogeneous graph network information and enhanced knowledge features were fused for sentiment classification prediction. Experimental results on public datasets （Twitter， Lap14， Rest15， Rest16） show that compared to models such as ISSK-GCN （Integrating Syntax and Sentiment Knowledge Graph Convolutional Network）， the proposed model achieves the best results， validating the effectiveness of the differentiated node modeling and knowledge enhancement strategies.

Multimodal sentiment analysis model for missing modalities under shared semantic conditions

Shang LIU, Zhaosen TANG, Hongyue LIU, Linfang DONG, Jin ZHOU

2026, 46(9):  2761-2768.  DOI: 10.11772/j.issn.1001-9081.2025080959

Asbtract ( )   HTML ( )   PDF (671KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Existing studies on multimodal sentiment analysis in modality missing scenarios often neglect inter-modal correlations when generating missing modalities， leading to semantic inconsistencies between restored and original data. Additionally， missing modality generation based on diffusion models brings large computational overhead. To address these issues， a Multimodal Sentiment Analysis Model for Missing Modalities under Shared Semantic Conditions （MM-SSC） was proposed. First， a shared latent space mapping module was designed to use the shared latent space of Vector Quantized Variational AutoEncoder （VQ-VAE） to capture multimodal distributions effectively， thereby ensuring cross-modal shared semantics. Second， a cross-modal consistency constraint method was proposed to learn mutual information within each modality's latent space， thereby promoting the refinement of semantic information between modalities and enhancing cross-modal consistency. Third， a missing modality reconstruction and alignment module was designed to reconstruct and refine missing modalities while reducing reconstruction computational overhead. Finally， a multimodal fusion and prediction module was introduced to fuse reconstructed and available modalities for consistent sentiment analysis. Experimental results demonstrate that under fixed modality missing conditions， compared with Incomplete Multimodality-Diffused emotion recognition （IMDer） model， the proposed model achieves average improvements of 0.4 and 0.7 percentage points in the F1 score and ACC7 on the CMU-MOSI dataset， respectively； on the CMU-MOSEI dataset， the F1 score and ACC7 are improved by 0.9 and 0.3 percentage points， respectively. Under random modality missing conditions， compared with IMDer model， the proposed model achieves an average improvement of 2.2 percentage points in both F1 score and ACC7 on the CMU-MOSI dataset； on the CMU-MOSEI dataset， the F1 score and ACC7 （accuracy for seven classes） are improved by 0.8 and 0.4 percentage points， respectively. It can be seen that MM-SSC can address multimodal sentiment analysis tasks under modality missing scenarios effectively.

Collaborative query optimization-based framework for long-tailed object detection

Jiangyan CHEN, Yandan WANG, Yihu LIU, Yinglong MA

2026, 46(9):  2769-2775.  DOI: 10.11772/j.issn.1001-9081.2025081018

Asbtract ( )   HTML ( )   PDF (1234KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the problem that significant performance degradation of Transformer-based object detection methods when facing long-tailed distributions， a collaborative query optimization-based single-stage end-to-end long-tailed object detection framework， named CQ-DETR （Collaborative Query optimization-based DEtection TRansformer）， was proposed. In the framework， a Layer-Adaptive Encoder Fusion （LAEF） module was designed to integrate multi-scale encoder features dynamically， so as to consider both high-level semantics and low-level details； a Feature-aware Query Generation （FQG） module was designed to generate content-aware queries dynamically from image features， thereby enhancing the representation capability of initial content queries for potential objects； a Category-Localization Joint-aware query Selection （CLJS） mechanism was proposed to achieve collaborative optimization of category coverage and localization accuracy. Experimental results indicate that CQ-DETR is superior to RichSem （Rich Semantics） method on the long-tailed object detection benchmark dataset LVIS v1.0， with the Average Precision （AP） and the AP of rare categories （AP_r） improved by 1.4 and 1.5 percentage points， respectively， verifying the effectiveness of the proposed framework in category-imbalanced scenarios； meanwhile， on the relatively balanced COCO 2017 dataset， compared with DINO （DETR with Improved deNoising anchOr box）， CQ-DETR has the AP improved by 1.1 percentage points， verifying the good generalization ability of this framework in general object detection scenarios.

Spatio-temporal modeling and hierarchical feature enhancement for person re-identification

Dingli YANG, Yuanfang WEI, Wenrui HU, Liyang KONG, Yinshan YU

2026, 46(9):  2776-2786.  DOI: 10.11772/j.issn.1001-9081.2025081053

Asbtract ( )   HTML ( )   PDF (2618KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the matching difficulties in person Re-IDentification （Re-ID） under complex scenes such as occlusion， viewpoint changes， and pose variations， an Re-ID algorithm based on spatio-temporal modeling and hierarchical feature enhancement was proposed. In the algorithm， a collaborative improvement of global consistency and local discriminability was achieved through a three-stage progressive feature optimization framework. First， after extracting appearance features by the backbone network， a dual-pooling temporal attention mechanism was introduced. In the mechanism， complementary information of sequence features was captured by combining global average pooling and temporal average pooling， and spatio-temporal dependency modeling was performed through channel and spatial interaction， so that the motion-related features were highlighted and the local information loss caused by occlusion was alleviated. Second， to address the problem of uneven distribution of human body part features， a flexible feature fusion module was constructed， so that the multi-part features were aggregated through learnable weights adaptively， occlusion noise was suppressed， and discriminative local features were enhanced， thereby obtaining the global and local hierarchical representations. Finally， a confidence correction network was designed before the classification layer， so that the identity prediction confidence distribution was optimized through residual learning， and the cross-camera retrieval accuracy was improved. The results of systematic evaluation on the Market-1501 public dataset show that the proposed algorithm achieves an mAP of 93.2% and a Rank-1 of 97.4%， which are increased by 3.8 and 1.7 percentage points compared with the baseline algorithm Body Part-Based ReID model （BPBReID_HR）. It can be seen that the proposed algorithm improves the Re-ID performance in complex scenes significantly by the combination of spatio-temporal modeling and hierarchical feature enhancement.

Lightweight human pose estimation framework based on wavelet attention mechanism with enhanced low frequency

Zhangjian JI, Siyuan WANG

2026, 46(9):  2787-2792.  DOI: 10.11772/j.issn.1001-9081.2025081009

Asbtract ( )   HTML ( )   PDF (568KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Human pose estimation is a critical task in computer vision and High-Resolution Networks （HRNets） are widely used due to their outstanding performance. However， improvements in model accuracy are often accompanied by a surge in parameter count and computational cost. To address this issue， an innovative lightweight improved module， Frequency Domain Cross Attention Convolution （FCAC）， was proposed. First， the basic convolutional module was reconstructed by fusing Wavelet Transform Convolutions （WTConvs） with Depthwise Separable Convolutions （DSCs） with WTConv's enhancing receptive field capability compensating for DSC's lack of contextual connections， thereby optimizing HRNet's performance while maintaining low complexity. Second， an attention mechanism was employed to guide low-frequency components to learn high-frequency components， and the learning results were aggregated dynamically through a dynamic channel reweighting mechanism. Furthermore， Squeeze-and-Excitation Network （SENet） was used to compensate for the current WTConv's learning of inter-channel feature relationships. Finally， FCAC was deployed at key nodes in the backbone network. Experimental results on the COCO dataset show that although the number of parameters of the improved HRNet-W32 model is increased by 22%， the model has the floating-point operations reduced by 7.0%， the mean Average Precision （mAP） improved by 3.5 percentage points， and the convergence ability greatly enhanced.

Video question answering method based on keyframes and summarization

Li HE, Zhiqiang LI, Yuhao SONG, Xiao WANG

2026, 46(9):  2793-2799.  DOI: 10.11772/j.issn.1001-9081.2025080995

Asbtract ( )   HTML ( )   PDF (1332KB) ( )  

Figures and Tables | References | Related Articles | Metrics

The video question answering task aims to understand video content deeply and answer natural language questions. Current methods based on Multimodal Large Language Models （MLLMs） struggle to model global semantics and temporal dependencies across scenes effectively due to dual constraints of context length and computational complexity， resulting in degraded model inference performance. To address this issue， a Scene-Driven Adaptive Keyframe Sampling （SD-AKS） method and a Question-Driven Video Summarization （QD-VS） method were proposed. First， an iterative video scene clustering based on K-means was introduced， and a Scene Separation Score （SSS） was designed， so as to evaluate feature differences between scenes quantitatively， thereby enhancing the accuracy and robustness of scene segmentation. Second， a problem semantic-guided adaptive keyframe sampling was designed to achieve dynamic keyframe selection based on quantified scene information density and fine coverage of information-dense regions. Furthermore， based on the keyframe method， Large Language Models （LLMs） were employed to summarize the video-related question set， so that the task-oriented textual summaries were generated to improve global model inference capability. Experimental results show that compared to LLaVA?Video， the best?performing baseline model， the proposed method improves prediction accuracy by 5.3 and 1.8 percentage points on the EgoSchema and NExT?QA datasets， respectively， which verifies its effectiveness in cross?dataset scenes.

Incremental semantic mapping system based on visual SLAM

Wei CUI, Ping ZHANG, Zhuoling XIAO, Zhuohang CHEN

2026, 46(9):  2800-2808.  DOI: 10.11772/j.issn.1001-9081.2025081035

Asbtract ( )   HTML ( )   PDF (1985KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Traditional semantic Simultaneous Localization and Mapping （SLAM） systems rely on training with full data， making it difficult to recognize novel categories in open environments. In addition， the systems only construct sparse maps， resulting in limited semantic representation capabilities and insufficient support for complex human-machine interaction. Therefore， an incremental semantic mapping system based on visual SLAM was proposed. In the system， a dense mapping thread was introduced into the ORB-SLAM3 （Oriented FAST and Rotated BRIEF SLAM 3.0） framework， and a newly designed semantic segmentation model， DAISS （Depth Attention Incremental Semantic Segmentation）， was integrated， so as to achieve simultaneous localization and semantic mapping. Among which， a dual-branch encoder and an attention distillation mechanism were adopted by the DAISS model to improve segmentation accuracy for old and new categories. Experimental results on NYU Depth V2 dataset show that this system achieves 61.0% and 47.5% in Mean Accuracy （MA） and Mean Intersection over Union （MIoU） metrics， representing improvements of 60.9% and 48.9%， respectively， over the SATS （Self-Attention Transfer for continual Semantic segmentation） model. These results validate its strong incremental learning capability and semantic perception performance.

Quaternion-based uncertain knowledge graph embedding

Renrui YIN, Guanfeng LI, Shizhuo WANG, Yaya JIANG

2026, 46(9):  2809-2819.  DOI: 10.11772/j.issn.1001-9081.2025081025

Asbtract ( )   HTML ( )   PDF (828KB) ( )  

Figures and Tables | References | Related Articles | Metrics

The existing methods have made preliminary progress in Uncertain Knowledge Graph Embedding （UKGE）， but still have multiple limitations， particularly in modeling complex relations. The UKGE framework can only model simple relation patterns such as symmetry. Although the complex-plane rotation model RotatE is simple and efficient， and can model multi-relation facts by complex rotations， its expressiveness is constrained by the two-dimensional complex space， making it difficult to model high-dimensional rotations， and it is prone to the “gimbal lock” problem， which limits further modeling of complex relations. To address these limitations， an uncertain knowledge graph embedding model based on quaternions， termed QUKGE， was proposed. In QUKGE， entities were represented as four-dimensional quaternion vectors， and relations were modeled as rotation operators in quaternion space. At the scoring function level， quaternion rotation was realized by utilizing the Hamilton product of entity and relation embeddings， thereby representing a variety of complex relation patterns systematically， including symmetry， asymmetry， inverse relations， commutativity， and non-commutativity. The three-dimensional rotation characteristics of quaternions effectively alleviate the “gimbal lock” problem in high-dimensional rotation and significantly enhance the capability to model complex relationships. To validate the effectiveness of QUKGE， five comparative experiments were conducted on three public uncertain knowledge graph datasets： CN15K， NL27K， and PPI5K. Experimental results demonstrate that on the NL27K dataset， QUKGE achieves a 2.9 percentage points improvement in F1 score and a 1.0 percentage point increase in accuracy compared to the MUKGE （Multiplex Uncertain Knowledge Graph Embedding） model on relational fact classification task， providing support for future research in UKGE.

Contract review framework based on dynamic rule generation and neuro-symbolic reasoning using large language models

Yaoran HUO, Huizhi LI, Dahe ZENG, Zhenyu TANG, Chenmei LIU, Yan WU, Xiao ZHANG, Siqin CHEN, Houhui YANG

2026, 46(9):  2820-2826.  DOI: 10.11772/j.issn.1001-9081.2025091093

Asbtract ( )   HTML ( )   PDF (656KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Automated contract review has limitations in adapting to diverse clauses dynamically and executing complex review rules accurately. Therefore， a Dynamic Rule generation and Neuro-Symbolic reasoning framework for contract Review （DRNS-Review） using Large Language Models （LLMs） was proposed. First， a dynamic rule generation method was designed， domain knowledge and contract context were integrated， and LLMs were used to generate natural language and symbolic review rules in real time. Then， a contract review mechanism based on neuro-symbolic reasoning was constructed to extract local facts automatically， rules were localized through symbolic matching， and explainable reasoning was completed by combining symbolic reduction with LLM semantic execution. Finally， reasoning outputs from multiple segments were aggregated， and a dual?track review report， comprising both structured and natural?language content， was generated. Experimental results demonstrate that DRNS-Review improves the risk identification F1 value by 13.0 and 8.3 percentage points， respectively， and improves the Risk Level Accuracy （RLA） by 13.3 and 8.7 percentage points， respectively， compared to baseline methods LLM-Direct and ManualRule-LLM. It can be seen that integrating dynamic rule generation with neuro-symbolic reasoning improves the accuracy and adaptability of contract review effectively， providing a generalizable technical pathway for intelligent compliance applications.

Data science and technology

Attribute-missing graph clustering model based on stacked joint optimization

Xiben LUO, Xiaoyun CHEN

2026, 46(9):  2827-2837.  DOI: 10.11772/j.issn.1001-9081.2025081008

Asbtract ( )   HTML ( )   PDF (974KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the inconsistency between attribute completion and clustering objectives in two-stage methods and the low efficiency of deep learning methods for attribute-missing graph clustering， an Attribute-Missing Graph Clustering model based on Stacked Joint Optimization （AMGC-SJO） was proposed. First， a Matrix Factorization-based Attribute-Missing Graph Clustering model （AMGC-MF） was constructed， which introduced an enhanced adjacency matrix to represent global node connection relations and employed graph regularization for attribute completion， and then a joint Non-negative Matrix Factorization （NMF） was applied to the completed the attribute matrix and enhanced adjacency matrix， so as to learn the node clustering degrees of membership. On this basis， an AMGC-SJO was further designed to update the attribute matrix and enhanced adjacency matrix dynamically during iteration， thereby enabling co-optimization of attribute completion and clustering tasks. Experimental results show that AMGC-SJO outperforms AMGC-MF on multiple clustering metrics， with better key metrics on five datasets. Compared to Attribute-Missing Graph Clustering （AMGC）， the proposed model achieves comparable clustering accuracy while reducing running time by at least 89.57% under the topological centrality-based missing mechanism. Furthermore， on the Cora dataset and attribute privacy missing mechanism， when the missing rate is increased from 10% to 90%， the clustering metrics of AMGC-SJO fluctuate within 2.08 percentage points， demonstrating strong robustness， whereas those of AMGC decline by over 30.88 percentage points. It can be seen that the proposed model provides an accurate， efficient， and robust solution for attribute-missing graph clustering.

Fair recommendation framework for large language models with sensitive attribute absence

Zhenhui GONG, Xiaoyu SHI, Yun LU, Yangcheng LIU, Mingsheng SHANG

2026, 46(9):  2838-2846.  DOI: 10.11772/j.issn.1001-9081.2025080969

Asbtract ( )   HTML ( )   PDF (793KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Large Language Models （LLMs） bring enhanced semantic understanding and personalized recommendation capabilities to recommender systems， however， they face significant challenges in user fairness in practical applications. The existing methods for LLM-based fair recommendation often rely on explicit sensitive attributes for constraints or reweighting， making them difficult to apply in scenarios where such attributes are unavailable due to privacy protection or inaccessibility. To address this issue， an FAIR recommendation framework for large language models with Sensitive Attribute Absence （FAIR-SAA） was proposed. In this framework， recommendation fairness was enhanced without accessing sensitive attributes through dynamic prompt optimization and adversarial reweighting. Specifically， in the first stage， a dynamic prompt optimization strategy was adopted to identify high-loss samples during the fine-tuning process， and these samples were utilized as contextual examples to mitigate stereotypical patterns； in the second stage， an adversarial reweighting mechanism was introduced to focus on underperforming regions of the model dynamically， thereby increasing the impact of underrepresented samples on model updates. Experimental results on three public datasets such as MovieLens-1M show that FAIR-SAA reduces the gender group Normalized Discounted Cumulative Gain NDCG@10 and Hit Ratio HR@10 by 74.47% and 61.76%， averagely， with a recommendation accuracy maintained comparable to baseline recommendation accuracy of the BI-step Grounding Paradigm for Recommendation （BIGRec）. When BIGRec is used as the base recommendation model， even when compared with Fairness-Aware Conformal Thresholding and Prompt EngineeRing （FACTER） framework， a fair recommendation method with full access to sensitive attributes， FAIR-SAA is competitive with 58.33% fairness metric. It can be seen that FAIR-SAA provides an effective solution to recommendation fairness problem in real-world privacy-preserving scenarios.

Force-directed layout visualization method for medical record knowledge graph presentation applications

Dongyang ZHOU, Jinlian DU, Xueyun JIN

2026, 46(9):  2847-2856.  DOI: 10.11772/j.issn.1001-9081.2025081037

Asbtract ( )   HTML ( )   PDF (6620KB) ( )  

Figures and Tables | References | Supplementary Material | Related Articles | Metrics

The force-directed layout method is a classical approach used for knowledge graph visualization， and it distributes nodes in the visualization space uniformly through attractive and repulsive forces between nodes. However， since visualization of domain-specific knowledge graphs is often closely related to domain applications， the classical force-directed layout method cannot meet the visualization requirements of domain-specific knowledge graphs. Therefore， an improved force-directed layout method was proposed by analyzing the graph structure， data characteristics， and visualization requirements of query applications of medical record knowledge graphs， with the aim of supporting visualization technology for medical record knowledge graphs in clinical assisted diagnosis and treatment applications. In the method， based on the classical force-directed layout model， attractive force between nodes of the same level and same type， repulsive force between nodes of the same level but different types， and repulsive force between non-leaf and non-adjacent nodes to maintain the spacing were introduced. As a result， nodes of the same level and the same type were grouped into the same region， nodes of the same level but different types were distributed into separate regions， and nodes at different levels were distributed with reasonable spacing， thereby reducing node crowding and edge crossings； so that the data’s type characteristics， quantity characteristics， and community characteristics were revealed by the final generated visualization layout clearly. Additionally， node type and the number of same-level， same-type nodes were taken as factors determining edge length between nodes， and the node importance was represented through edge length， thereby further improving the rationality of the visualization layout and enhancing semantic visualization. Experimental results show that in terms of semantic visualization capability in vertical domains， the proposed method achieves an average improvement of 165.5% in the mean Node-Type Spatial Discriminability （NTSD） compared with the FR （Fruchterman-Reingold） method. As the visualization results， this method has better performance for electronic medical record knowledge graph visualization.

Cyber security

Cross-domain handover authentication scheme for VANETs with vehicle type verification

Yunlong WANG, Yali LIU, Dongdong CHEN, Yi ZHOU, Xiaohui PANG

2026, 46(9):  2857-2866.  DOI: 10.11772/j.issn.1001-9081.2025080988

Asbtract ( )   HTML ( )   PDF (1662KB) ( )  

Figures and Tables | References | Related Articles | Metrics

In order to solve the problems in cross-domain authentication for Vehicular Ad-hoc NETwork （VANET）， which include the inability to distinguish vehicle types， low efficiency in authentication， and insufficient security， a Cross-domain Handover Authentication Scheme for VANETs with Vehicle Type Verification （CHAS-VTV） was proposed. In specific， in order to solve key escrow and complex certificate management problems while improving cross-domain authentication efficiency， a certificateless cross-domain authentication mechanism based on the certificateless public key cryptography system was proposed； concerning the classified cross-domain authentication of vehicles， a dynamic accumulator-based mechanism was proposed to store and verify vehicle types； in order to solve the problem of repeated authentication among vehicles and multiple RoadSide Units （RSUs）， a Hash-based Message Authentication Code （HMAC）-based handover authentication mechanism was proposed， which enabled efficient handover authentication and key negotiation among vehicles and multiple RSUs. Performance analysis results show that compared to the existing cross-domain handover authentication schemes for VANET， CHAS-VTV reduces computational cost by at least 3.87% and communication cost by at least 14.28% in cross-domain authentication， while in handover authentication， it reduces computational cost by at least 18.09% and communication cost by at least 14.48%. Security analysis results show that CHAS-VTV scheme can resist various malicious attacks effectively， which includes public key substitution attacks， private key leakage attacks， and forgery attacks.

ADS-B attack detection method based on improved DeepLabV3+

Lixia XIE, Chaoyue CAO, Hongyu YANG

2026, 46(9):  2867-2876.  DOI: 10.11772/j.issn.1001-9081.2025081040

Asbtract ( )   HTML ( )   PDF (1113KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the problems that the communication messages of Automatic Dependent Surveillance-Broadcast （ADS-B） system lack encryption and identity authentication mechanisms and are vulnerable to spoofing attacks such as replay and trajectory forgery， but traditional detecting models cannot identify anomalies in the system， an ADS-B attack detection method based on an improved DeepLabV3+ model was proposed. First， the original tabular ADS-B data were reconstructed into grayscale images through grayscale mapping， thereby enabling the visualization of spatio-temporal features. Then， to overcome the weak feature extraction capability and blurred segmentation boundaries of DeepLabV3+ model when processing ADS-B image data， a Haar Wavelet Downsampling （HWD） module was used to replace the max-pooling operation in the ResNet101 backbone network， thereby preserving key features such as trajectory anomalies and local perturbations during dimensionality reduction effectively. Finally， a Simple parameter-free Attention Mechanism （SimAM） was introduced after the ADS-B low-level feature map to enhance the representation of local spatial structures， and a Fusion Convolutional Block Attention Module （F_CBAM） was incorporated after the Atrous Spatial Pyramid Pooling （ASPP） module to highlight critical semantic regions through the parallel interaction of channel and spatial attentions. Experimental results on an ADS-B dataset show that compared with the baseline DeepLabV3+ model， the proposed method achieves a 6.73 percentage points improvement in mean Intersection over Union （mIoU）. In complex attack scenarios， the proposed method has the F1-scores for Denial of Service （DoS） attacks and random noise reached 96.07% and 92.55%， respectively， which are significantly higher than those of mainstream models such as U-Net （U-shaped convolutional Network）， SegFormer （Segmentation Transformer）， and K-Net （Kernel Network）， providing a new approach for accurate detection of ADS-B attacks.

DCSFFuzzer： dual-channel semantic feature fusion-based fuzz testing method for industrial control protocols

Kan HE, Hongfeng MA, Xuejun ZONG, Hongyan SHI, Lian LIAN, Bowei NING

2026, 46(9):  2877-2888.  DOI: 10.11772/j.issn.1001-9081.2025081026

Asbtract ( )   HTML ( )   PDF (1285KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the limits of the existing fuzz testing techniques in deep semantic modeling and feature representation of Industrial Control Protocols （ICPs）， leading to high test case redundancy and low acceptance rates， an ICP fuzz testing method based on dual-channel semantic feature fusion， DCSFFuzzer， was proposed. In the method， dual-channel parallel architecture was adopted to perform multi-level semantic modeling for ICPs， and a Generative Adversarial Network （GAN） was used to generate diverse test cases. First， global semantic features were extracted by capturing global dependencies in protocol sequences with a Transformer encoder， and local field features were extracted by capturing relationships among adjacent bytes or fields with a multi-scale one-dimensional Convolutional Neural Network （CNN） encoder. Second， the extracted multi-level semantic features were fused by a Gated Recurrent Unit （GRU） adaptively to enhance key semantic feature representation， thereby increasing the acceptance rate of test cases. Finally， a “relativistic discriminator” training strategy was introduced in the GAN to enhance diversity of test cases by comparing relative authenticity of real and generated samples to alleviate mode collapse. Based on the above methods， the fuzz testing framework DCSFFuzzer was designed， and experiments were conducted on three ICPs： Modbus/TCP， S7comm， and Ethernet/IP. Experimental results show that compared with five models： TDRFuzzer， MLFRFuzzer， WGANFuzzer， GANFuzzer， and PeachFuzzer， DCSFFuzzer has the Test Case Acceptance Rate （TCAR） and Diversity of Generated Data （DGD） improved， and the Test Target Anomaly Rate （TTAR） increased by 0.016， 0.024， 0.069， 0.130， and 0.172 percentage points， respectively， validating the method’s effectiveness. It can be seen that by fusing multi-level semantic features， DCSFFuzzer covers more ICP vulnerabilities， improving the security detection capability of industrial control systems.

End‑to‑end robust video watermarking against screen‑recapturing attacks based on spatio-temporal feature enhancement network

Yimin ZHOU, Wenhui FU, Jie LUO, Juan WANG

2026, 46(9):  2889-2897.  DOI: 10.11772/j.issn.1001-9081.2025080998

Asbtract ( )   HTML ( )   PDF (1717KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Recently， with the advancement of multi-source capturing devices and sensing technologies， screen recapturing has become a primary channel for video information leakage during dissemination. The screen-recapturing process causes irreversible physical damage to video watermark information， making it difficult to trace the source of information leakage， which poses a significant challenge in the field of video forensics. The existing anti-screen-recapturing video watermarking techniques are still in the preliminary research stage， and are mainly relying on handcrafted frequency-domain features， which struggle to resist cross-device screen-recapturing attacks effectively. Therefore， an end-to-end anti-screen-recapturing video watermarking method was proposed on the basis of a spatio-temporal feature enhancement network. First， to enrich watermark embedding features， a spatio-temporal feature enhancement network module was designed by integrating an optical flow estimation network and a high-pass filter， thereby optimizing the watermark information embedding strategy. Then， by introducing key factors of screen-recapturing channel noise to simulate screen-recapturing distortion， an anti-screen-recapturing attack noise layer was designed to enhance cross-device robustness against such attacks. Experimental results in real-world screen-recapturing attack scenarios demonstrate that the proposed method achieves a decoding accuracy of over 95%， with both robustness and image quality outperforming the mainstream solutions such as Robust Inherent Video Information Embedding （RIVIE） and Automatic， Robust and Blind Video waterMarking resisting camera recording （ARB-VM）.

Advanced computing

Heterogeneous multi-agent reinforcement learning enabled co-optimization of UAV 3D obstacle avoidance and edge computing

Guanliang CHEN, Yi LIU, Yi YU

2026, 46(9):  2898-2909.  DOI: 10.11772/j.issn.1001-9081.2025080956

Asbtract ( )   HTML ( )   PDF (1700KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Severe challenges in terms of real-time processing and low-energy transmission brought by the development of Internet of Things （IoT） and the proliferation of mobile terminal devices are face by compute-intensive tasks. Particularly in multi-Unmanned Aerial Vehicle （UAV） -assisted Mobile Edge Computing （MEC） scenarios， communication links are constrained by obstacle blockages and UAV trajectories in complex 3D environments， and latency and energy consumption pressures are further exacerbated. For the scenario of multiple UAVs providing computational offloading services to ground users， an optimization model was established to minimize the weighted sum of the system's maximum task completion latency and total energy consumption， so as to optimize the users' discrete offloading decisions and the UAVs' continuous 3D trajectories jointly. To solve the problem of mixed （discrete-continuous） action space and strong decision coupling， a heterogeneous multi-agent algorithm UOUM （User Offloading and UAV Mobility co-optimization） was proposed. In the algorithm， under a heterogeneous multi-agent deep reinforcement learning framework， dedicated network architectures were designed for the two types of heterogeneous agents： users and UAVs. And differential reward mechanism was introduced to quantify marginal contributions of the agents， thereby solving the multi-agent credit allocation problem. Concurrently， an Artificial Potential Field （APF） was innovatively integrated as a differentiable physical constraint into the agent learning framework， so as to ensure safe obstacle avoidance for the UAVs. Simulation results show that compared to three benchmark methods （Only User Offloading optimization （OUO）， Only UAV Trajectory optimization （OUT）， and a standard heterogeneous multi-agent reinforcement learning （H-MARL） using only a global reward mechanism））， UOUM has advantages in various scenarios with different numbers of users， UAVs， and obstacle densities. Compared with H-MARL， UOUM has the final convergence reward improved by approximately 28.6% on average， and achieves strong environmental adaptability in terms of latency control， energy optimization， and safe obstacle avoidance.

Trajectory-aware multi-agent cooperative task offloading method for vehicular edge computing

Chenyang WANG, Xiaoyu SHI, Jie GAN, Mingsheng SHANG

2026, 46(9):  2910-2919.  DOI: 10.11772/j.issn.1001-9081.2025080966

Asbtract ( )   HTML ( )   PDF (1209KB) ( )  

Figures and Tables | References | Related Articles | Metrics

In Vehicular Edge Computing （VEC） environments， the high mobility and uneven distribution of vehicles often lead to cross-domain task offloading failures and load imbalance of RoadSide Units （RSUs）， thereby degrading system performance significantly. The existing studies mainly focus on load balance or single-step offloading decisions， while failing to address the communication uncertainties caused by vehicle mobility and the requirement for flexible task scheduling across RSUs， which limits offloading efficiency. Therefore， a Trajectory-Aware Collaborative Multi-Agent Reinforcement Learning （TAC-MARL） method was proposed. First， foresighted mobility awareness for task scheduling was realized through introducing a Patch Time Series Transformer （PatchTST） -based multi-step vehicle trajectory prediction module to guide task scheduling. Second， a vehicle-edge-cloud collaborative task offloading framework was constructed， in which vehicles and RSUs were modeled as agent groups， cooperative decision-making was performed by the employment of a reinforcement learning paradigm — Centralized Training and Decentralized Execution （CTDE）， and a partially reward-decoupled multi-agent policy optimization method was designed to improve cooperation efficiency and system stability. Simulations under varying task densities， latency constraints， and network topologies were conducted. The proposed method consistently outperforms baseline algorithms such as Independent Proximal Policy Optimization （IPPO）， Multi-Agent Deep Deterministic Policy Gradient （MADDPG）， and Multi-Agent Proximal Policy Optimization （MAPPO） in task completion rate and task completion latency， verifying the efficiency， robustness， and practical potential of the proposed method in dynamic and complex VEC environments.

Adaptive constrained multi-objective evolutionary algorithm guided by effective information

Yuanxin LI, Hecheng LI, Xiaojing HAN

2026, 46(9):  2920-2930.  DOI: 10.11772/j.issn.1001-9081.2025080993

Asbtract ( )   HTML ( )   PDF (917KB) ( )  

Figures and Tables | References | Related Articles | Metrics

For Multi-objective Optimization Problem （MOP）， the goal is to obtain a well-distributed set of Pareto optimal solutions. When the variables are constrained， the Pareto solutions are typically located in different feasible regions. For large-scale and tightly constrained optimization problems， the existing evolutionary algorithms have difficulty in balancing diversity and convergence of individuals， resulting in the search population may not being able to cross multiple infeasible regions and then converging prematurely. To address this issue， an effective Information-guided Adaptive Constrained Multi-objective Evolutionary Algorithm （IACMEA） was proposed. First， a co-evolutionary mode of main and auxiliary populations was adopted： the main population mainly searched within the feasible regions and converged to the Pareto fronts， and the auxiliary population retained infeasible solutions with small objective function values through environmental selection based on the constraint threshold， thereby maintaining population diversity. Second， to improve the quality of individuals in the populations， a mating pool selection strategy was designed on the basis of the feasible solution ratio. Experimental results on the CF， LIRCMOP， and DASCMOP benchmark test sets show that， comparing IACMEA with six similar algorithms including CCMO （Coevolutionary Constrained Multi-objective Optimization framework）， IACMEA achieves 23 best Inverted Generational Distance （IGD） values and 22 best HyperVolume （HV） values out of 33 problems. It can be observed that IACMEA demonstrates certain advantages in handling nonlinear Constrained Multi-objective Optimization Problems （CMOPs）.

Variable entropy-based warning propagation algorithm for solving minimum cut problem

Simin HU, Xiaofeng WANG, Hongsheng DING, Jiahuan SONG, Xiaona SUO, Dong YAN

2026, 46(9):  2931-2937.  DOI: 10.11772/j.issn.1001-9081.2025070904

Asbtract ( )   HTML ( )   PDF (768KB) ( )  

Figures and Tables | References | Related Articles | Metrics

The Minimum Cut Problem （MCP） is a classic NP （Nondeterministic Polynomial） -hard problem with broad applications in image segmentation and network reliability analysis. Warning Propagation （WP） algorithm， a factor graph-based inference method， has shown good scalability and structural adaptability when solving combinatorial optimization problems. However， its effectiveness on MCP is limited by path selection， factor graph construction， and convergence. Therefore， an Entropy-based Warning Propagation （EWP） algorithm was proposed to solve MCP. First， a Hidden Markov Model （HMM） -driven jump-point mechanism was introduced to select core propagation regions. At the same time， a batch-mapping factor graph transformation way was employed to reduce graph structural dimensionality effectively. Finally， a propagation freezing strategy that combined marginal probability and variable entropy was adopted to accelerate convergence. Experimental results demonstrate that EWP achieves good performance on graphs with different sizes， and especially has significant potential in large-scale graph optimization problems.

Computer software technology

Agile enhancement development framework driven by human-multi-agent collaboration

Mengjiao YU, Min JIN, Xinhua LIU

2026, 46(9):  2938-2947.  DOI: 10.11772/j.issn.1001-9081.2025081043

Asbtract ( )   HTML ( )   PDF (1177KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Agile development， with its iterative and rapid delivery characteristics， has become a mainstream software development mode. However， it still faces challenges such as unsustainable evolution， difficulty in knowledge retention， and collaboration issues caused by skill disparities. With the rise of Large Language Models （LLMs） and agent-based technologies， an agile enhancement development framework driven by human-multi-agent collaboration was proposed with LangChain as the collaborative core and Retrieval-Augmented Generation （RAG） as the knowledge support. The framework was optimized through three core mechanisms： an architecture design Agent was introduced to generate three differentiated architectural schemes — conservative （adapting to current requirements but requiring long-term refactoring）， innovative （supporting complex， future-oriented implementations）， and balanced （mediating between current and future needs） — which were evaluated using a seven-dimensional quantitative assessment （including scalability） to assist human architects in decision-making， thereby addressing the challenges of insufficient modeling of technical debt sources and predicting architectural evolution. Development artifacts were automatically generated through agents and predefined templates. Role-based modeling， grounded in a knowledge base， defined the functional boundaries and execution standards of agents， precisely matching tasks with capabilities to reduce collaboration bottlenecks. Experimental results on a portal website project demonstrate that， compared with traditional Scrum methods， the proposed framework reduces development cycle time by 29.6%， decreases defect density by 46.9%， and improves user satisfaction by 16.5%， effectively validating its engineering value and promotion potential in enhancing automation and knowledge management， while ensuring delivery efficiency and system quality.

Construction method of software defect prediction dataset with explainability

Yun BIAN, Haiquan WANG, Yi CHEN, Zhe CUI

2026, 46(9):  2948-2958.  DOI: 10.11772/j.issn.1001-9081.2025080987

Asbtract ( )   HTML ( )   PDF (859KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Software defect prediction often lacks explainable information， such as defect localization， defect explanation， and repair suggestions， making the prediction results difficult to be applied in actual development. To address this problem， an explainable method for constructing a software defect prediction dataset was proposed on the basis of context engineering and Large Language Models （LLMs）， and HandPick， the first multi-programming-language dataset for software defect prediction， was released. First， the TriCogVuln-LLM （LLM-enhanced Triple Cognitive chain for multilingual code Vulnerability mining with common weakness enumeration knowledge） method was designed on the basis of software engineering principles and prior defect knowledge， so as to guide LLMs to complete function description generation， Common Weakness Enumeration （CWE） defect predictions， and repair suggestion generation step by step. Second， a consensus voting mechanism was designed to form an optimal generative model pool for defect prediction， thereby further improving the quality and diversity of the generated data. Finally， the HandPick dataset， covering four mainstream programming languages， was constructed through consensus-driven automated data generation. Downstream task validation results show that the Qwen2.5-14B-HandPick model， fine-tuned on the HandPick dataset， achieves significant improvements compared to baseline models on an independent and public test set， with gains of 19.29， 21.26， 24.11， and 18.30 percentage points in precision， recall， F1-score， and accuracy， respectively. This confirms that the HandPick dataset， generated by the TriCogVuln-LLM method， enhances model's defect identification and analysis capabilities significantly.

Multimedia computing and computer simulation

Superpixel segmentation algorithm based on inverted feature pyramid network

He SUN, Guanghui YAN, Xiaohong JIA

2026, 46(9):  2959-2967.  DOI: 10.11772/j.issn.1001-9081.2025070895

Asbtract ( )   HTML ( )   PDF (3080KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Superpixels exhibit excellent image representation capability and high computational efficiency， which are widely applied in subsequent computer vision tasks. However， existing superpixel segmentation algorithms fail to fully integrate semantic and spatial information， resulting in insufficient spatial coherence and semantic consistency of the generated superpixels， as well as blurred and misclassified details when processing images with complex scenes. To address these problems， this paper proposed a Superpixel Segmentation algorithm based on the Inverted Feature Pyramid Network （SSIFPNet）. Through the cascaded coupling of an encoder and an Inverted Feature Pyramid Network （IFPN） decoder， the proposed algorithm better fused end-to-end multi-scale semantic information and spatial detail information. Meanwhile， a perceptual boundary-driven composite loss function was designed to further improve the semantic consistency of superpixels. Comparative experiments on the BSDS500， NYUv2， and KITTI datasets were conducted to compare SSIFPNet with state-of-the-art algorithms， including the Simple Linear Iterative Clustering superpixel segmentation algorithm （SLIC）， Superpixel Sampling Network （SSN）， and Superpixel segmentation algorithm with fully Convolutional neural Network （SCN）. Experimental results show that SSIFPNet achieves superior performance across all datasets. Specifically， the achievable segmentation accuracy （ASA） and Boundary Precision （BP） of SSIFPNet reach 0.972 0 and 0.130 6 on the BSDS500 dataset， 0.948 2 and 0.199 9 on the NYUv2 dataset， and 0.965 5 and 0.147 0， along with a COmpactness （CO） of 0.372 0， on the KITTI dataset. It is demonstrated that the superpixels generated by SSIFPNet possess favorable spatial coherence and semantic consistency and achieve superior performance in processing detailed texture regions， which fully demonstrates the superiority and generalizability of SSIFPNet.

Remote sensing small object detection with dynamic perception and cross-modulation

Hua YAO, Gaoming YANG, Xuelian LI, Kaixuan LU

2026, 46(9):  2968-2976.  DOI: 10.11772/j.issn.1001-9081.2025080985

Asbtract ( )   HTML ( )   PDF (1245KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the challenges of severe background interference and insufficient multi-scale feature extraction and fusion for small object detection in remote sensing images， which lead to degraded detection accuracy， a Dynamically Perceptive and Cross-Modulated Network （DPCMNet） was proposed. First， to enhance the model’s perception ability to fine-grained objects， the multi-level Hybrid Feature Aggregation （HFA） module was designed to strengthen the response intensity of small object regions dynamically， thereby enriching the textural and structural features effectively. Second， the Dynamic Modulation Fusion （DMF） module was proposed to introduce a joint channel and spatial modulation mechanism， so as to enable adaptive feature interaction， thereby alleviating information loss during multi-scale feature fusion and improving localization and recognition capabilities for small objects. Finally， based on the DMF module， the Dynamic Modulation Fusion Pyramid （DMFP） was constructed to achieve progressive feature fusion and enhancement through cross-scale association between high-level semantic information and low-level detail compensation. Experimental results demonstrate that DPCMNet achieves significant performance improvements on the Unicorn Small Object Dataset （USOD） and VEhicle Detection in Aerial Imagery （VEDAI） remote sensing small object datasets， with Average Precision （AP） increased by 2.9 and 3.1 percentage points， respectively， compared to the baseline model YOLOv11n.

Small-object defect detection method for printed circuit boards based on dual-branch feature extraction

Yufeng OU, Chia-Hung WANG

2026, 46(9):  2977-2986.  DOI: 10.11772/j.issn.1001-9081.2025081034

Asbtract ( )   HTML ( )   PDF (1251KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the challenges of detecting small-object defects on Printed Circuit Board （PCB） surfaces， such as low contrast， irregular shapes， and susceptibility to background interference， a dual-branch feature detection model named Dual branch feature Extraction-Spatial attention Detection （DES-Det） was proposed. First， a dual-branch collaboration was used to extract global contextual and local detailed features， thereby enhancing the discriminative capability for low-pixel defects. Second， a deformable attention mechanism was introduced to achieve adaptive feature focusing， thereby improving the localization accuracy for blurred edges and irregular defects. At the same time， a detection head based on spatial attention and depthwise separable convolution was designed to reduce model complexity significantly while maintaining detection performance. Experimental results on two public PCB defect datasets demonstrate that DES-Det achieves precisions of 93.6% and 98.4%， respectively， representing improvements of 5.7 and 1.4 percentage points， respectively， compared to the baseline model YOLOv8， with the number of parameters reduced by 26.7%. Cross-scenario generalization experimental results further validate the method's effectiveness and robustness in complex industrial environments.

REL-YOLO： lightweight road waterlogging detection network integrating reflection perception features

Aoyun YAN, Linglong ZHU

2026, 46(9):  2987-2995.  DOI: 10.11772/j.issn.1001-9081.2025080989

Asbtract ( )   HTML ( )   PDF (1777KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Road waterlogging is regarded as a widespread traffic safety hazard that can easily lead to vehicle loss of control and then trigger traffic accidents. Accurate identification of road waterlogging is limited by challenges such as complex road surface textures， variable lighting conditions， and the diverse morphology of puddle itself. To address these issues， REL-YOLO （Reflection-aware Enhanced Lightweight YOLO）， a lightweight road waterlogging detection network integrating reflection perception features， was proposed on the basis of YOLOv11n. First， the Reflection-Aware Feature Enhancement （RAFE） module was designed to fuse shallow texture and deep semantic features effectively， and specular reflection interference was suppressed through an adaptive noise weighting mechanism. Second， the C3k2_Enhanced module was designed by combining grouped convolution with a lightweight channel attention mechanism， so that the number of parameters was reduced while the capture capability of fine water flow features was enhanced. Finally， a Lightweight Shared Detail Enhanced Convolutional Detection head （Detect_LSDECD） was introduced， and multi-scale feature fusion was optimized through shared convolution and learnable scale parameters. Experimental results show that compared with the original YOLOv11n model on the Road waterlogging dataset， REL-YOLO has recall improved by 3.2 percentage points， mAP@50 improved by 2.0 percentage points， the number of model parameters reduced by 13.1%， and floating-point operations reduced from 6.3 GFLOPs to 4.3 GFLOPs. While maintaining high real-time performance， REL-YOLO achieves a balance between accuracy and efficiency.

Glomerulonephritis classification based on hierarchical aggregation and prompt enhancement

Guibin ZHANG, Yun MI, Yanmeng LU, Jian GENG, Zhitao ZHOU, Lei CAO

2026, 46(9):  2996-3004.  DOI: 10.11772/j.issn.1001-9081.2025080954

Asbtract ( )   HTML ( )   PDF (1710KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Accurately distinguishing GlomeruloNephritis （GN） subtypes is crucial for clinical decision-making， however， GN automated diagnosis faces challenges in modeling the intrinsic relationships between non-consecutive Transmission Electron Microscopy （TEM） images within a bag and utilizing the fine-grained expert knowledge in pathology reports effectively. Therefore， an image-text multimodal model， CLIP-HaRe （CLIP framework with HTA and RME）， was proposed， comprising a Hierarchical Transformer Aggregation （HTA） module and a Report Memory Enhancement （RME） module. In HTA module， a hierarchical aggregation strategy “sub-image-instance-bag” was utilized， in which a self-attention mechanism was employed to model the semantic relationships among non-consecutive TEM images， and max-pooling was used to extract key lesion features. In RME module， the pathology report memory banks were constructed using modern Hopfield network， the class prompts generated by a Large Language Model （LLM） were employed for query， and fine-grained textual knowledge was integrated from the reports， thereby enhancing the disease-specific representation capability of the prompts. Experimental results in the disease classification task on a constructed dataset of three GN subtypes show that the proposed model achieves an accuracy of 0.773 7， a precision of 0.781 6， a recall of 0.772 3， an F1 score of 0.772 7， and an Area Under the receiver operating Characteristic curve （AUC） of 0.930 8. These metrics represent improvements of 11.01， 10.68， 10.69， 11.04， and 9.85 percentage points， respectively， over a baseline method Linear-Probe. CLIP-HaRe enables the automated classification of GN subtypes by fusing image and text information， with its significant performance enhancement over the baseline， validating its potential for application in assisting clinical diagnosis.

Lung cancer image classification method via bidirectional collaborative distillation and multi-instance learning

Jinchang LI, Wei JIA, Xuefen ZHAO

2026, 46(9):  3005-3016.  DOI: 10.11772/j.issn.1001-9081.2025080960

Asbtract ( )   HTML ( )   PDF (1067KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the issue in the existing Whole Slide Image （WSI） classification for lung cancer， where Multi-Instance Learning （MIL） methods based on knowledge distillation architecture struggle to handle the mixture of hard and easy instances effectively， leading to insufficient model learning capability and imbalanced performance， a lung cancer whole slide image classification method based on Bidirectional Collaborative Distillation and Multi-Instance Learning （BCD-MIL） was proposed by incorporating the concept of multi-task learning. First， a Bidirectional Collaborative Distillation Framework （BCDF） comprising multiple student models and a teacher model was designed， so as to satisfy the differentiated learning requirements for easy and hard pathological instances in lung cancer histopathological images， and by distributing different tasks to multiple student models， the model's learning capacity was enhanced while solving performance imbalance problem. Concurrently， classification performance was improved through a virtuous cycle with the teacher model. Second， a Multi-Angle Instance Mining Parallel architecture （MA-IMP） was designed to match the heterogeneous differences in lung cancer pathological features across dimensions such as cell morphology and tissue texture， and by conducting instance mining from multiple perspectives， the mining bias caused by a single perspective was avoided. Finally， a Dynamic Stage-Aware Exponential Moving Average distillation （DSA-EMA） algorithm was proposed to optimize weight update of the teacher model， and improve training efficiency and model performance based on the stage characteristics of large-scale instance training of lung cancer histopathological images through adjusting distillation parameters in the training phase dynamically. Experimental results show that compared to the MIL framework with Masked Hard Instance Mining （MHIM-MIL） method， which is also based on knowledge distillation and instance mining， BCD-MIL achieves improvements of 1.92， 2.37， 3.43， and 1.75 percentage points in Area Under Curve （AUC）， accuracy， F1-Score， and recall， respectively， on The Cancer Genome Atlas （TCGA） dataset， and improvements of 1.11， 6.29， 6.94， and 12.85 percentage points in four metrics， respectively， on the Clinical Proteomic Tumor Analysis Consortium （CPTAC） dataset； validating the effectiveness of the proposed method. Furthermore， the lightweight distillation architecture reduces model parameter size and inference time， thereby enabling efficient deployment while ensuring performance gains， and providing a reliable basis for lung cancer WSI classification.

Frontier and comprehensive applications

Hybrid CNN-Transformer multi-task learning method with dynamic task weighting for radar signal feature recognition

Hao WU, Yulu HUANG, Xuan FENG, Qiong ZHANG, Xiaobo ZHANG

2026, 46(9):  3017-3024.  DOI: 10.11772/j.issn.1001-9081.2026020161

Asbtract ( )   HTML ( )   PDF (1672KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the problems in the existing methods for radar signal feature recognition： the poor performance in low Signal-to-Noise Ratio （SNR） environments and the difficulty in handling multiple tasks of classification and parameter estimation simultaneously， a hybrid multi-task learning method fusing Convolutional Neural Network （CNN） and Transformer was proposed. In the method， multi-scale local feature extraction was integrated with global dependency modelling. To further achieve balanced joint optimization， a dynamic task weighting strategy fusing Bayesian uncertainty estimation and task-difficulty awareness was introduced to adjust the loss weights adaptively. Besides， an attention visualization module was developed to enhance model interpretability. Experimental results on the RadChar-Baseline dataset show that compared with baseline models such as IQST （IQ Signal Transformer）， CNN1D （One-dimensional Convolutional Neural Network）， CNN2D （Two-Dimensional Convolutional Neural Network）， ResNet1D （One-Dimensional Residual Network）， and Long Short-Term Memory （LSTM） network， the proposed method achieves a composite score of 0.931； within the SNR range of -20 dB to 20 dB， it has an average composite score of 0.937 with a standard deviation of only 0.108. It can be seen that the proposed method significantly outperforms the comparative methods in both classification and regression tasks， especially under challenging low SNR conditions.

CNN-Transformer encoder-based model for depressive state recognition with multidimensional EEG feature fusion

Siyu DONG, Xin JIA, Mi ZHANG, Xinwen CHEN, Xinying SUN, Fangxiao CHENG

2026, 46(9):  3025-3034.  DOI: 10.11772/j.issn.1001-9081.2025070908

Asbtract ( )   HTML ( )   PDF (903KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Considering the challenges of insufficient feature fusion and limited classification granularity in ElectroEncephaloGram （EEG） -based depression recognition task， a CNN-Transformer Encoder-based model for Depressive State Recognition （CTE-DSR） that integrated multidimensional EEG features was proposed. First， One-Dimensional Convolutional Neural Network （1D-CNN） was used to extract local spatial sequential features from EEG. Then， Transformer encoder module was introduced to model global correlations between leads. Simultaneously， Multi-Layer Perceptron （MLP） was used to embed depression-related non-sequential features into a high-dimensional space. Finally， the two features were fused to implement fine-grained classification of depression states through a fully connected layer. Besides， a loss weighting strategy was employed to improve the model's ability to identify minority classes. Experimental results show that compared with other five machine learning models （such as Random Forest （RF） and Adaptive Boosting （AdaBoost）） and two combination models （CNN-LSTM and CNN-RNN）， CTE-DSR achieves at least 23.0% and 6.7% higher recall and Area Under Curve （AUC） in binary classification， respectively， and at least 15.6% and 1.0% higher F1-score and macro-average AUC in ternary classification， respectively. It can be seen that the proposed model has performance advantages in identifying depressive states and can be used to assist in early screening and relapse risk assessment of depression.

EEG decoding model based on dual-branch representation fusion and cross-modal alignment

Xiaocui XU, Bo LI, Yutong ZOU

2026, 46(9):  3035-3042.  DOI: 10.11772/j.issn.1001-9081.2025081016

Asbtract ( )   HTML ( )   PDF (658KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Concerning the problem that the existing text-oriented ElectroEncephaloGraphy （EEG） decoding methods focus on global modeling while neglecting inter-channel local correlations and failing to align EEG and text representations precisely， an EEG decoding model based on dual-branch representation fusion and cross-modal alignment was proposed. In the model， a parallel dual-branch architecture was adopted to obtain high-quality representations of EEG signals： in the spatio-temporal branch， signal short-term dependencies， neighboring-channel spatial correlations， and cross-step long-range relationships were captured layer by layer through Bidirectional Long Short-Term Memory （Bi-LSTM） network， depthwise separable convolution， and gated axial self-attention mechanism； in the context fusion branch， based on a multi-layer Transformer encoder， cross-attention was used to fuse the two representations for complementary integration. Besides， to reduce semantic discrepancies between modalities， a joint loss combining triplet loss and covariance alignment was introduced， thereby constraining vector alignment by the geometric distance and second-order statistical properties between paired EEG and text representations. Experimental results on the ZuCo dataset show that the proposed model improves BLEU-1 （BiLingual Evaluation Understudy-1） by about 1.16 percentage points compared with the baseline EEG2Text， demonstrating its effectiveness in text-oriented EEG decoding tasks.

Continuous non-invasive blood pressure prediction method based on multi-scale Transformer with enhanced graph neural network

Changqing JI, Ziying ZHANG, Zumin WANG

2026, 46(9):  3043-3053.  DOI: 10.11772/j.issn.1001-9081.2025070894

Asbtract ( )   HTML ( )   PDF (1256KB) ( )  

Figures and Tables | References | Related Articles | Metrics

Hypertension is a major risk factor for CardioVascular Disease （CVD）， and continuous blood pressure monitoring is crucial for CVD prevention. To address the limitations of the existing models in multi-scale feature representation， long-term dependency modeling， and inter-task information interaction， a data-driven blood pressure prediction model named Multi-scale Transformer with Enhanced Graph neural network （MTEG） was proposed to integrate a multi-scale convolution， an enhanced Transformer encoder， a channel attention mechanism， and a graph attention-based multi-task learning strategy to model local and global features of PhotoPlethysmoGraphy （PPG） signals jointly. Specifically， a multi-scale feature extraction module was designed to capture local patterns across different temporal scales in PPG signals， while a Transformer encoder with Rotary Position Embedding （RoPE） and relative position bias was employed to model long-term dependencies in blood pressure variations， a feature-aware enhancement module was introduced to strengthen the response of channels related to blood pressure changes， and a multi-task feature aggregation module was constructed to achieve information collaboration among the prediction tasks of Systolic Blood Pressure （SBP）， Diastolic Blood Pressure （DBP）， and Mean Arterial Pressure （MAP）. Experimental results show that MTEG achieves the Mean Absolute Error （MAE） of 4.92， 2.68， and 2.59 mmHg in SBP， DBP， and MAP predictions， respectively. Predictions for DBP and MAP comply with the standards of the Association for the Advancement of Medical Instrumentation （AAMI） and achieve grade A of the British Hypertension Society （BHS） standard， while SBP prediction achieves grade B of the BHS standard， verifying the reliability and clinical application potential of the proposed model.

Oil reservoir sweet spot identification method based on Shapelet-Transformer

Mingzhe LIU, Xiaoqin XIONG, Shixun BAI, Renbao ZHAO, Ziwei WANG, Faqiang GOU, Liang XU

2026, 46(9):  3054-3064.  DOI: 10.11772/j.issn.1001-9081.2025081033

Asbtract ( )   HTML ( )   PDF (2172KB) ( )  

Figures and Tables | References | Related Articles | Metrics

To address the issues that oil reservoir geological data are complex to represent and the existing methods struggle to fully extract the features of oil reservoir sweet spots， an oil reservoir sweet spot identification method based on Shapelet-Transformer was proposed. In this method， well log data were re-represented as a multi-variable depth sequence and a network model integrating local and global features was constructed to achieve accurate identification of oil reservoir sweet spots. First， a local feature learning unit was proposed， and the unique local well log response features of sweet spots were captured through the Shapelet extraction and selection mechanism. Second， a global feature learning mechanism was constructed to mine the global trend and parameter correlation of well log sequences， thereby providing global geological features. Finally， a fusion and classification module was designed to output the identification results of oil reservoir sweet spots by fusing local and global features. Experimental results show that the average identification accuracy of this method on the well log data of Ma 2 well block is 84.62 %， which is at least 1.26 percentage points higher than that of Fully Convolutional Neural Network （FCNN）， BiLSTM-FCNN （Bidirectional Long Short-Term Memory and FCNN hybrid model）， Temporal Graph Convolutional Network （T-GCN）， Disaggregated Multi-Scale Time Series decomposition model （DisMS-TS） and other benchmark models， which verifies the effectiveness of this method.