Most Read articles

    Published in last 1 year |  In last 2 years |  In last 3 years |  All

    Published in last 1 year
    Please wait a minute...
    For Selected: Toggle Thumbnails
    Journal of Frontiers of Computer Science and Technology    2026, 20 (1): 0-0.  
    Related Articles | Metrics
    Abstract (4593)
    PDF (101)
    Automatic Prompt Engineering Technology for Large Language Models: a Survey
    BA Zezhi, ZHANG Hui, XIE Zhenghan, ZUO Xiaodong, HOU Jianwei
    Journal of Frontiers of Computer Science and Technology    2025, 19 (12): 3131-3152.   DOI: 10.3778/j.issn.1673-9418.2502027
    Prompt engineering (PE) based on prompt learning is crucial for improving the technical accessibility of LLMs and accelerating their adoption, diffusion, and application development. Compared with traditional PE, which heavily relies on the domain knowledge and experience of prompt designers and is less adaptable to tasks with large prompt spaces, automatic prompt engineering (APE) can generate or optimize prompts in an automatic or semi-automatic way. This enables the exploration of large-scale prompt combinations and enhances the stability of prompt generation through automated optimization techniques. However, there is currently a lack of systematic reviews on APE, which hinders subsequent researchers from quickly grasping the state of the field. Therefore, this paper keeps up with the latest research developments, systematically reviews the implementation forms of automated prompt engineering, and proposes future research directions. Based on the trade-offs in logical reasoning and performance orientation in the implementation of a APE, this paper categorizes it into four main types: APE based on chain-of-thought, APE based on machine learning models, APE based on evolutionary algorithms and plug-and-play auto-prompt systems. Subsequently, this paper conducts a comprehensive evaluation of APE techniques, constructing a theoretical explanatory framework for their working principles and assessing the applicability and limitations of each implementation form. Finally, this paper looks ahead to the development trends of APE in multimodal large models, advanced reasoning models and AI-Agents.
    Reference | Related Articles | Metrics
    Abstract (909)
    PDF (2519)
    Survey of Model Compression for Large Language Model
    GUO Jinyang, HE Changyi, YANG Ge, LIU Xianglong
    Journal of Frontiers of Computer Science and Technology    2026, 20 (1): 1-20.   DOI: 10.3778/j.issn.1673-9418.2504069
    Large language models (LLMs) have attracted considerable attention in recent years due to their strong cognitive capabilities and widespread applications in various fields. However, their tremendous demand for computation and memory makes it difficult to deploy them in resource-constrained scenarios. Model compression and acceleration techniques have thus emerged as critical approaches to reduce computational complexity and memory usage while maintaining model performance. This paper presents a comprehensive survey of recent advances in LLM compression and acceleration methods, aiming to grasp the current development status and future trends of the entire field, promote the advancement of LLM compression and acceleration technologies, and facilitate their application and implementation in both industry and academia. It begins by outlining the challenges LLMs face in terms of computational and storage overhead. Then, it categorizes and reviews the main technical approaches, including model pruning, quantization, knowledge distillation, and low-rank decomposition, highlighting their core principles, representative methods, and cutting-edge developments. In addition, the paper provides a detailed discussion on evaluation metrics such as inference latency, accuracy retention, and deployment cost, establishing a multidimensional evaluation framework. Finally, it explores the promising future directions of LLM compression methods, aiming to guide future research and industrial deployment of compressed LLMs.
    Reference | Related Articles | Metrics
    Abstract (707)
    PDF (2755)
    Overview of Multimodal Generation for Recommender Systems
    ZHANG Rui, BIAN Zhipeng
    Journal of Frontiers of Computer Science and Technology    2025, 19 (12): 3224-3242.   DOI: 10.3778/j.issn.1673-9418.2511039
    With the rapid development of large language models and multimodal generative techniques, recommender systems are moving from matching existing content to generating personalized content, which motivates the emerging task of personalized multimodal generation. This task focuses on producing user aligned text, image, audio, or video content based on historical behaviors and target instructions so that the generated outputs can be directly used as recommendation candidates or display materials. Although recent studies have shown promising results in several modalities, existing research remains fragmented and lacks a unified definition, technical formulation, and methodological structure. To address this gap, this survey provides a systematic overview of personalized multimodal generation in recommender systems. It formalizes a three component modeling relationship consisting of preference capture, target content, and personalized generation, and clearly defines the task scope as generating multimodal outputs that serve recommendation purposes (such as cover images, news headlines, audio and video clips) rather than general text to image or open domain dialogue generation. It then constructs a unified framework covering preference and target modeling, preference injection and generator architectures, and optimization strategies for personalized outputs, and summarizes representative techniques across image, text, audio, and cross-modal generation together with typical application scenarios. In addition, it analyzes the limitations of existing evaluation metrics, highlights their insufficiency in measuring personalization and recommendation effectiveness, and discusses the challenges brought by large multimodal models in terms of model adaptation, inference efficiency, and system safety. Finally, this survey outlines future research directions to support the continued development of personalized multimodal generation in recommender systems.
    Reference | Related Articles | Metrics
    Abstract (589)
    PDF (279)
    TaDPT: Multi-task Learning Fusion Framework for Multi-behavior Sequential Recommendation
    LIU Zhuang, WANG Kunhao, GUO Xiaoyang, JIANG Yujing, SHANG Jiaming
    Journal of Frontiers of Computer Science and Technology    2026, 20 (7): 2009-2026.   DOI: 10.3778/j.issn.1673-9418.2508020
    With the rapid evolution of the digital economy and the diversification of user interactions, recommendation systems in e-commerce platforms face new modeling challenges. User behaviors have shifted from single clicks to multi-stage interactions, including browsing, adding to cart, and purchasing, which exhibit strong heterogeneity and temporal dependencies. Existing approaches often focus on single behavior modeling or coarse-grained fusion at the output layer, failing to simultaneously capture long-term dependencies and short-term patterns, thus limiting recommendation performance. To address these issues, this paper proposes a multi-task learning fusion framework for multi-behavior sequential recommendation, termed type-aware dual-path transformer (TaDPT). Built upon a Transformer backbone, TaDPT introduces a dual-path architecture: a self-attention path to capture global temporal dependencies and a frequency-domain filtering path (FMLP) to enhance local feature representation. A type-aware SoftSelect gating mechanism is designed to adaptively fuse the two paths, while a behavior-type adaptive positional bias is incorporated into the attention mechanism to dynamically adjust temporal decay across different behaviors, and more precisely capture the evolutionary characteristics of multiple behaviors. Extensive experiments on the public OTTO dataset demonstrate that TaDPT outperforms state-of-the-art methods on key metrics such as AUC and HR@20, while maintaining a favorable balance between inference efficiency and scalability, providing a practical and theoretically grounded solution for multi-behavior recommendation tasks.
    Reference | Related Articles | Metrics
    Abstract (541)
    PDF (399)
    Survey of Deep Learning-Based UAV Single Object Tracking
    CHEN Long, SHI Lei, LI Zhihui, DING Meng, PAN Yilun
    Journal of Frontiers of Computer Science and Technology    2026, 20 (1): 40-65.   DOI: 10.3778/j.issn.1673-9418.2506046
    Deep learning-based UAV (unmanned aerial vehicle) single object tracking has emerged as a critical research area in computer vision, aiming to accurately track designated targets in aerial video sequences. UAV tracking presents unique challenges, including drastic perspective changes, variable target scales, and computational constraints. This survey systematically categorizes recent methods into three technical approaches: traditional Siamese networks, CNN-Transformer hybrid architectures, and full Transformer methods, focusing on advances from 2022 to 2025. This paper proposes innovative sub-classifcation frameworks, including: module replacement, feature post-fusion, and collaborative modeling for CNN-Transformer hybrid architectures; static computation, hybrid mechanisms, and dynamic computation for single-stream Transformer methods. These frameworks reveal the evolution from performance-oriented to efficiency-performance balanced optimization. Comprehensive evaluations on UAV123, DTB70, UAVDT, and VisDrone2018 datasets validate the advantages and limitations of different approaches. This paper identifies key challenges with future directions and engineering deployment guidance.
    Reference | Related Articles | Metrics
    Abstract (536)
    PDF (2518)
    Multi-agent Collaborative Framework for Audit Case Qualification and Regulation Recommendation
    XU Chao, LIU Zishuo, ZHOU Liyun, ZHU Haoran, HUANG Jiajia
    Journal of Frontiers of Computer Science and Technology    2026, 20 (1): 280-290.   DOI: 10.3778/j.issn.1673-9418.2503038
    Addressing the technical challenges of multidimensional feature analysis and dynamic regulation alignment in audit case qualification, this paper develops an audit regulation recommendation system based on multi-agent collaboration. Conventional large language models exhibit three limitations in regulatory recommendation tasks: semantic discrepancies between legal texts and audit problem representations, insufficient knowledge currency due to dynamic regulatory updates, and inadequate multi-level reasoning capabilities under single retrieval strategies. To resolve these issues, a collaborative architecture is proposed, integrating specialized agents for problem extraction, logical reasoning, and pseudo-case generation. The task decomposition mechanism parses complex audit cases into parallelizable subproblem spaces. Methodologically, the framework innovatively combines instruction fine-tuning with retrieval-augmented generation (RAG) technology, establishing a three-tier regulatory knowledge base covering national statutes, organizational policies, and internal control protocols. A dynamic agent-based retrieval strategy is incorporated to enhance precision. Experimental results demonstrate 30.56% direct clause matching accuracy in national regulation recommendations, outperforming ChatGPT-4 (13.89%) by 116%. Semantic evaluation metrics achieve 71.19% BERTScore and 20.20% Rouge-L. For multi-level regulatory recommendations, the system attains 63.91% accuracy, 26.50% BERTScore, and 27.80% Rouge-L, which exceeds the baseline model by at least 15.2%. The findings reveal that task decomposition through agent collaboration effectively disentangles multidimensional audit features, while the hierarchical knowledge architecture improves recommendation accuracy and provides interpretable regulatory reasoning paths. The modular design supports dynamic expansion of jurisdiction-specific regulation databases, demonstrating significant engineering applicability.
    Reference | Related Articles | Metrics
    Abstract (481)
    PDF (171)
    Research Survey on Emotion Recognition Large Language Models
    WU Ao, WANG Hailong, LIU Lin, SHI Wentao
    Journal of Frontiers of Computer Science and Technology    2026, 20 (3): 625-649.   DOI: 10.3778/j.issn.1673-9418.2509014
    The rapid development of large language models offers a new paradigm for emotion recognition, demonstrating significant advantages over conventional methods in complex scenario understanding, zero/few-shot generalization, and multimodal collaborative representation. This paper provides a systematic review of large-scale emotion recognition models. It first summarizes advances in unimodal emotion recognition across text, speech, visual, and physiological signals. The study then focuses on multimodal approaches, categorizing them into unified encoder architectures, hierarchical fusion architectures, and generative model architectures based on fusion strategies, and compares their design principles, key techniques, and applicable scenarios. Further analysis highlights improvements in recognition accuracy, generalization capability, and multimodal fusion, underscoring the important role and potential of large language models in this field. Finally, this paper identifies challenges such as insufficient temporal emotion modeling and cross-cultural cognitive biases, and suggests future directions including enhancing long-range temporal modeling and establishing multimodal cultural emotion benchmarks.
    Reference | Related Articles | Metrics
    Abstract (438)
    PDF (2443)
    Visual Mamba: Structure, Practice, and Prospects
    ZHANG Xin, ZHI Min, Sarula, Arimuzha
    Journal of Frontiers of Computer Science and Technology    2026, 20 (1): 66-78.   DOI: 10.3778/j.issn.1673-9418.2503061
    Traditional convolutional neural networks (CNNs) struggle to model global features due to their limited receptive field. Though vision Transformers (ViTs) possess the advantage of sequence modeling, they face the issue of quadratic computational complexity, posing severe computational challenges for image processing. In response, researchers have begun exploring new architectures that combine efficient computation with global perception capabilities. The visual Mamba model, based on state space models (SSMs), enables global context modeling under linear computational complexity while retaining sequence modeling capabilities, marking a new stage in vision modeling based on state space models. This paper elaborates on the basic framework of the visual Mamba block, including its dual residual structure composed of residual modules, 2D selective scan (SS2D) modules, and feed-forward networks (FFN). It analyzes the working mechanisms of cross-scanning, S6 block processing, and cross-fusion within the SS2D module. The visual Mamba model is explored from three aspects: scanning methods, stacking methods, and hybrid architectures. Scanning methods include sequential scanning and dynamic scanning, with a comparative analysis of the advantages and disadvantages of different scanning strategies. Stacking methods are categorized into serial Mamba, parallel Mamba, U-shaped Mamba, and graph Mamba, with a detailed analysis on the network construction logic of each stacking structure and its adaptability in multi-scale feature extraction and long-range dependency modeling. Hybrid architectures focus on fusion forms with CNNs, Transformers, and attention mechanisms, including single-module fusion and multi-module collaborative architectures, along with an analysis of the strengths and weaknesses of each model. Through analysis, it is pointed out that the visual Mamba model overcomes the local perception limitation of CNNs and the quadratic computational complexity of Transformers. It outperforms mainstream backbone architectures in visual tasks and demonstrates tremendous potential to become a fundamental visual backbone.
    Reference | Related Articles | Metrics
    Abstract (422)
    PDF (2343)
    Survey on Time-Series Data Anomaly Detection Methods
    XU Jianqiu, HU Wei, JIN Duosi, GAO Yunjun, ZHAO Xujian
    Journal of Frontiers of Computer Science and Technology    2026, 20 (6): 1517-1544.   DOI: 10.3778/j.issn.1673-9418.2510014
    Recently, time-series data anomaly detection has received widespread attention from academia and industry, with crucial applications in service monitoring, IoT (Internet of things) surveillance, medical diagnosis, and cybersecurity. Ongoing research efforts in this domain have led to a substantial accumulation of research outcomes. To this end, this paper provides a comprehensive review of anomaly detection in time-series data and proposes a unified taxonomy, which divides all methods into three categories: shallow, deep learning-based, and model selection, with the latter two categories further subdivided. This paper also offers a detailed description of the principles and characteristics for each method. Among these three categories, shallow methods do not rely on deep learning structures but instead leverage domain knowledge from outlier detection, data mining, and classical machine learning for lightweight anomaly detection. Deep learning-based methods, based on principles of prediction, reconstruction, or contrastive learning, use the representation learning capabilities of various neural network architectures to achieve anomaly detection. Model selection methods, based on model selection or model ensembling principles, integrate multiple models to achieve the domain generalization that individual models lack. Commonly used datasets, evaluation metrics, benchmarks and open-source tools for time-series data anomaly detection are analyzed and summarized. A system architecture for time-series data anomaly detection is proposed, and the complete end-to-end process from time-series data acquisition to anomaly detection application is outlined. The major challenges in time-series data anomaly detection are summarized, along with insights into future research trends.
    Reference | Related Articles | Metrics
    Abstract (416)
    PDF (868)
    Survey of Graph Anomaly Detection Driven by Graph Neural Networks
    XU Dengbin, YUAN Lining, WU Peichen, LIU Zhao
    Journal of Frontiers of Computer Science and Technology    2026, 20 (5): 1207-1240.   DOI: 10.3778/j.issn.1673-9418.2505028
    Graph anomaly detection has been widely applied in network security, financial transaction monitoring, and social network analysis, aiming to identify anomalous nodes, edges, or subgraphs within graph databases. Leveraging their ability to model complex structural patterns, graph neural networks have emerged as a key approach in this domain. However, existing surveys exhibit gaps in coverage and classification coarseness. To address these shortcomings, this survey systematically reviews graph neural network-based anomaly detection research from three perspectives. It provides precise definitions of different anomaly types in graph data, and introduces a refined classification framework by grouping models according to graph data type (static vs. dynamic) and anomaly level (node, edge, subgraph, graph-level, and multi-level). It traces the development and representative architectures of various methods through the lenses of joint embedding and anomaly detection optimization, contrastive learning, and autoencoder-based reconstruction. It compiles mainstream public datasets, evaluates the performance of representative models, and delves into major challenges such as robustness, efficiency, cross-domain transferability, and lightweight deployment. Based on this analysis, the survey outlines future research directions for multi-level anomaly detection, cross-domain adaptation, and lightweight scenarios.
    Reference | Related Articles | Metrics
    Abstract (404)
    PDF (1292)
    DEPA-YOLO:Drone-Based Small Object Detection Model
    LIU Chenjie, LIU Wei, YANG Wendi, WANG Cheng
    Journal of Frontiers of Computer Science and Technology    2026, 20 (1): 266-279.   DOI: 10.3778/j.issn.1673-9418.2506003
    In response to the challenges of small object detection from the UAV perspective, including dense target distribution, complex background interference, and insufficient feature resolution, this paper proposes an improved model based on YOLOv10, named DEPA-YOLO. A DCMB (dynamic cross-modal bottleneck) is designed, which integrates dynamic weight allocation and multi-pattern feature mixing strategies to enhance the joint modeling capability of shallow local textures and high-level global semantics, thereby improving the feature representation of small objects. A HFEP (hierarchical feature enhancement pyramid) structure is proposed, which combines SPDConv (space-to-depth convolution) with ECFI (efficient cross-branch feature integration) multi-branch dynamic fusion to preserve fine details across different scales and efficiently transmit multi-level semantic information, significantly alleviating the information loss of traditional feature pyramids in dense scenarios. A CPCA (channel prior convolutional attention) attention mechanism is embedded into the backbone network to guide the model to focus on salient target regions, reducing false positives and missed detections caused by complex backgrounds. A WIoU (wise intersection over union) regression loss is adopted, introducing gradient regulation and boundary penalty terms to improve the stability and robustness of small object localization. Ablation experiments on the VisDrone2019 dataset demonstrate that DEPA-YOLO improves precision, mAP50, and mAP50:95 by 4.7, 4.3, and 2.5 percentage points, respectively, compared with the YOLOv10 baseline. Furthermore, generalization experiments on the DOTA dataset show additional improvements of 7.9, 2.0, and 1.2 percentage points. The proposed DEPA-YOLO model significantly enhances the detection accuracy of small objects from UAV perspectives while maintaining efficiency, providing an effective solution for UAV-based small object detection.
    Reference | Related Articles | Metrics
    Abstract (395)
    PDF (228)
    RCRQ: Multi-agent Information Extraction Framework for Traffic Accident Texts
    SHI Danyu, WANG Jingsheng, CONG Haozhe, WANG Erna, LIU Mingshuai
    Journal of Frontiers of Computer Science and Technology    2026, 20 (5): 1403-1416.   DOI: 10.3778/j.issn.1673-9418.2508069
    Building on the police traffic authority??s reliance on textual evidence for handling and regulating autonomous-vehicle (AV) incidents, this paper addresses extraction challenges posed by heterogeneous report genres, frequent cross-sentence coreference, disordered timelines, and missing causal statements. This paper proposes a multi-agent entity-event-relation structured information-extraction framework that treats large language models (LLMs) as agents and produces inter-pretable, high-confidence outputs through multi-stage processing, to support knowledge-graph construction and rule mining. For the constructed Accident-Benchmark dataset, this paper generates gold annotations via a dual-model collaboration with human-in-the-loop (HITL) verification. Then, this paper uses progressive prompt screening to select the best performer as the single-agent baseline. This paper designs an RCRQ MA framework that enhances extraction through staged reasoning, targeted error diagnosis, iterative repair, and consistency checking, and introduces a joint distillation method that leverages both reasoning trajectories and final outputs to further reduce inference cost. Experiments show that, relative to the single-agent baseline, the RCRQ MA framework yields substantial gains: on DeepSeek-V3, F1 for entities, event triggers, and event arguments improves by 1.89, 7.32, and 5.75 percentage points, respectively; on GLM-4-9B, the human-evaluation average score for relation extraction increases by 0.4 on a five-point scale. Both models outperform alternative baselines, validating the effectiveness of the proposed approach.
    Reference | Related Articles | Metrics
    Abstract (383)
    PDF (168)
    Survey on Event Extraction: from Traditional Methods to Large Language Models
    CAO Jiangjiang, DING Cangfeng, MA Lerong, SHI Dongyan, NING Qinwei
    Journal of Frontiers of Computer Science and Technology    2026, 20 (5): 1294-1322.   DOI: 10.3778/j.issn.1673-9418.2507007
    Event extraction, as one of the core tasks in natural language processing, aims to automatically identify event elements (such as trigger words, event types, arguments, and their roles) from unstructured text and transform them into structured representations. This technology has broad application prospects and significant practical value in scenarios such as public opinion monitoring, financial analysis, and disaster early warning. Given its critical position in theoretical research and practical applications, as well as the technical challenges it faces, this paper systematically reviews the development trajectory of event extraction technology. First, it clarifies the basic definitions and structural framework of this task, analyzing the main differences in information coverage, semantic integration capabilities, and processing strategies between sentence-level and document-level event extraction. Next, from the perspective of extraction frameworks, this paper summarizes three main frameworks, including pipeline-based, joint-based, and end-to-end, and systematically analyzes their operation processes, representative methods, and applicable scenarios. Then, it outlines the technological paradigm evolution from traditional rule-based and feature-engineering methods to deep learning approaches, and further to emerging methods based on pre-trained language models and large language models, analyzing the advantages, limitations, and development potential of each method in practical applications. Meanwhile, it reviews commonly used datasets and evaluation metrics, providing standardized references for method comparison and performance evaluation. Finally, based on the current research status, this paper identifies the key challenges in event extraction, such as generalization ability, structural expression, and semantic integration, and provides an outlook on future research directions.
    Reference | Related Articles | Metrics
    Abstract (372)
    PDF (1194)
    Multi-object Tracking Method for UAVs Based on Adaptive Illumination Enhancement
    WU Yuanxun, WANG Yabin, ZHANG Dawei, ZHENG Zhonglong
    Journal of Frontiers of Computer Science and Technology    2026, 20 (2): 452-465.   DOI: 10.3778/j.issn.1673-9418.2509060
    Multi-object tracking (MOT) holds significant promise across various applications, including intelligent transportation and video surveillance. However, unmanned aerial vehicle (UAV) scenarios present unique challenges, such as prevalent low-illumination conditions, rapid maneuvers, and dense, small-scale targets. To address these issues, this paper proposes an illumination-adaptive multi-object tracking framework (ILMOT) centered on the principles of “on-demand enhancement, lightweight gating, and geometry-constrained association”. The goal is to enhance the robustness of both detection and identity association in low-light environments. ILMOT employs a training-free IllumiGuard gating unit for robust low-light determination. This unit utilizes sRGB luminance statistics, sky region cropping, and bi-directional histogram trimming to reliably identify low-light conditions, triggering enhancement only when necessary. This on-demand approach significantly reduces computational overhead. When triggered, a pre-trained Retinex-Net is employed for a process of “decomposition-illumination enhancement-reflectance denoising-multiplicative reconstruction”. This approach suppresses noise amplification and halo artifacts while boosting visibility in dark areas, ensuring the quality of input for subsequent tracking modules. The tracker is based on DepthMOT’s three-branch architecture, leveraging EUCB (efficient up-convolution block) to improve depth estimation quality and integrating anchor-free detection with camera pose estimation to enhance feature representation and tracking accuracy. An unscented Kalman filter (UKF) is utilized for camera motion compensation and uncertainty propagation, effectively addressing target drift caused by rapid UAV movements. Depth information-based cascaded matching reduces ID switches and trajectory fragmentation in crowded scenes, enhancing tracking robustness. Experiments on the VisDrone2019 and UAVDT datasets validate the effectiveness of ILMOT. On VisDrone2019, ILMOT achieves significant improvements in HOTA, MOTA, and IDF1 metrics, while also effectively reducing ID switch counts, demonstrating its superior performance in low-light conditions. On UAVDT, which primarily features normal lighting conditions, ILMOT also exhibits robust performance. Ablation studies on a specifically created LMOT dataset demonstrate the effectiveness of low-light selective enhancer in low-light environments. The results indicate that the combined design of illumination-adaptive enhancement and geometry-constrained association can effectively improve the overall accuracy and identity consistency of multi-object tracking in complex UAV scenarios involving low light and high target density.
    Reference | Related Articles | Metrics
    Abstract (367)
    PDF (184)
    Supply Chain Finance Credit Risk Assessment Based on LLM Enhanced Graph Structure
    LIU Ying, LIU Shuang, LU Yu, SUN Nan, CHEN Jiamao
    Journal of Frontiers of Computer Science and Technology    2026, 20 (2): 522-532.   DOI: 10.3778/j.issn.1673-9418.2507013
    In the context of intensified global economic integration, the rising complexity of corporate supply chains has increasingly amplified the demand for financial services. Big data-driven credit risk assessment for supply chain finance can realize fast and accurate risk pricing. However, there are still three major limitations: (1) Most of the current assessment methods rely only on structured data, ignoring the potential risk signals in unstructured data; (2) Most of the modeling methods based on graph neural networks consider the complex relationships and conduction mechanisms between entity-level enterprises, ignoring the importance of attribute features; (3) The multilayered aggregation of graph neural networks leads to the problem of over-smoothing, which makes the model insufficiently identify the differences between different enterprises. To solve these problems, this paper proposes a credit risk assessment framework LGAT (large language model-enhanced graph attention network) for supply chain finance with LLM augmented graph structure. Firstly, this paper integrates the management discussion and analysis (MD&A) content in the annual reports of listed companies to achieve the expansion of data modality and enhancement of information dimension. Secondly, this paper makes lightweight fine-tuning of the LLM to reduce the computational cost of model training and inference while extracting the emotional features. Lastly, this paper conducts the fine-grained information aggregation from the perspective of feature dimensions and the perspective of enterprise neighbors, and introduces a graphical attention neural network with gated residual connections to control the information flow between the original input features and the transformed features to avoid excessive smoothing of the node features. By constructing a supply chain finance dataset for comparative analysis, the experimental results show that LGAT significantly outperforms the traditional model in the risk assessment task.
    Reference | Related Articles | Metrics
    Abstract (358)
    PDF (169)
    MAFS-YOLO: Improved YOLO11-Based Algorithm for Small Object Detection from UAV Perspective
    LIU Yansong, GAO Shuhui, WEI Dingding
    Journal of Frontiers of Computer Science and Technology    2026, 20 (6): 1782-1794.   DOI: 10.3778/j.issn.1673-9418.2511054
    To address the critical challenges in unmanned aerial vehicle (UAV) aerial imagery such as difficulty in detecting small objects, significant background interference, and multi-scale object variations, an enhanced object detection model named multi-branch attention and Fourier-based SPPF YOLO (MAFS-YOLO) is proposed based on the YOLO11 architecture. During the feature extraction stage, multi-branch attention network with wavelet enhancement module (MANet) is introduced, which decomposes input features into four frequency bands via wavelet transform and integrates parallel convolutional pathways to achieve multi-scale fusion of spatial details and frequency-domain structures. This design notably strengthens the perception of local textures and contours of small objects. In the backbone network, a Fourier-based spatial pyramid pooling fast (FourierSPPF) module is designed, leveraging fast Fourier transform to map features into the frequency domain. Through low-, medium-, and high-frequency masks, information from different frequency bands is separated and reconstructed before being restored to the spatial domain via inverse transform, thereby enhancing the ability of the model to capture global image structures and contextual relationships. A shape-decoupled IoU (SDIoU) regression strategy is employed in the loss function, where the matching of bounding box width and height is decoupled into two independent loss terms, improving the adaptability of the model to targets of varying scales. Experimental results on the VisDrone-DET2019 dataset demonstrate that MAFS-YOLO outperforms the baseline YOLO11 model, achieving improvements of 3.2, 1.9, and 1.6 percentage points in precision, mAP@0.5, and mAP@0.5:0.95, respectively, while reducing the number of parameters by approximately 13%. In generalization experiments on the TinyPerson dataset, corresponding gains of 3.5, 1.7, and 1.5 percentage points are observed. The experiments confirm that MAFS-YOLO significantly enhances the recognition accuracy of small objects while maintaining detection efficiency, providing a reliable technical pathway for object detection in complex UAV scenarios.
    Reference | Related Articles | Metrics
    Abstract (356)
    PDF (233)
    Construction of Industry Chain Knowledge Graph via Large Language Model-Based Verification Enhancement
    ZHENG Aoze, ZHANG Kunli, LI Yunlong, WANG Ying, YUAN Songrui, WU Pengcheng, JIA Yuxiang, ZAN Hongying
    Journal of Frontiers of Computer Science and Technology    2025, 19 (12): 3328-3339.   DOI: 10.3778/j.issn.1673-9418.2505073
    With the deepening of industrial chain analysis and the increasing demand for information processing, efficiently extracting supply-demand relationships between enterprises and industries from massive text data and constructing industrial chain knowledge graphs have become a critical research issue. To address this challenge, this paper proposes a knowledge graph construction method for industrial chains enhanced by large language model verification. At the conceptual level, this paper designs a hierarchical architecture for graph construction from dual perspectives of “industry-enterprise” to establish mapping relationships between enterprises and industrial sectors. At the data level, structured data undergoes knowledge extraction through large language models to develop an industrial chain domain corpus for entity and relation annotation (corpus for entity and relation annotation in industrial chain domain, CERIC). For unstructured data, this paper implements a verification-augmented relation triple extraction LLM (VRTE-LLM), which employs the CERIC corpus to fine-tune the large language model, thereby enhancing domain-specific knowledge comprehension and information extraction capabilities. The framework systematically rectifies completeness errors and deviation errors in relational triples by integrating contextual analysis with task-specific requirements, while applying a pre-defined rule base for triple validation to significantly improve the accuracy of enterprise information identification from unstructured texts. Experimental results demonstrate that the framework achieves 80.9% accuracy in entity recognition and 83.9% accuracy in relation extraction. The constructed industrial chain knowledge graph comprises over 39627 triples, covering four major industrial domains and knowledge representations of more than 70 enterprises, with comprehensive mapping across critical upstream and downstream chain segments. This methodology provides technical support for quantitative analysis and trend prediction in industrial chain research.
    Reference | Related Articles | Metrics
    Abstract (354)
    PDF (123)
    Review of YOLO Algorithm Research for 2D Medical Image Detection
    GUO Zhen, LIU Jing, QIU Dawei, LI Yuhao
    Journal of Frontiers of Computer Science and Technology    2026, 20 (1): 79-98.   DOI: 10.3778/j.issn.1673-9418.2502055
    In recent years, the breakthrough development of artificial intelligence technology has promoted a paradigm change in the interdisciplinary field of medicine and engineering, among which the object detection algorithm based on deep learning has shown significant advantages in medical image analysis. As a typical representative of the single-stage detection framework, the YOLO (you only look once) series algorithms have demonstrated unique advantages of high real-time, strong generalization ability and precise positioning in the field of medical image analysis through the “end-to-end” detection paradigm, and have gradually become the mainstream research methods for lesion detection, cell recognition and other tasks. The research of YOLO improved algorithm for medical object detection is sorted out. Firstly, based on the dimension of algorithm architecture innovation, the core evolution path of 12 generations of basic algorithms from YOLOv1 to YOLOv11 is sorted out, and the improvement breakthroughs, advantages and limitations, and medical scene performance of each version of YOLO are compared and analyzed. Secondly, the classic open-source datasets in the field of medical object detection are summarized, and the commonly used evaluation indicators in object detection are expounded. At the same time, the literature research on the use of YOLO improved algorithm in the detection of cervical cells, blood cells, pulmonary nodules and diabetic retinopathy in 2D medical images is reviewed, and different improved methods are comprehensively compared and analyzed. Finally, this paper summarizes the medical scenarios corresponding to different improvement ideas of YOLO, and discusses the challenges and future development directions in this field.
    Reference | Related Articles | Metrics
    Abstract (351)
    PDF (2199)
    LDD-YOLO: Improved Lightweight Dense Pedestrian Detection Algorithm Based on YOLOv8
    YANG Di, ZHANG Xilong, WANG Peng
    Journal of Frontiers of Computer Science and Technology    2026, 20 (1): 251-265.   DOI: 10.3778/j.issn.1673-9418.2504085
    To address the problems of missed detection and false detection caused by occlusion and scale variation in the current pedestrian detection algorithms in dense scenes, as well as the high computational complexity of the model, this paper proposes a YOLOv8-based lightweight dense pedestrian detection algorithm (LDD-YOLO), aiming to balance detection efficiency and accuracy. Specifically, a reparameterized efficient layer aggregation network (RELAN) is designed by integrating reparameterized convolutions and multi-branch structures, which respectively enhance feature representation during training and inference efficiency during deployment. Subsequently, a spatial pyramid pooling module with large-scale kernel attention (SPPF-LSKA) is introduced. This module combines depthwise large-kernel convolution operations to enlarge the receptive field, improve the ability to capture dense targets, and effectively suppress background interference. Moreover, to address the limitation of YOLOv8 in insufficiently capturing both local and global information during feature processing, an improved multi-scale feature fusion module, feature-focused diffusion module (FFDM) is proposed. By integrating multi-scale feature information, FFDM enhances the model’s feature representation capabilities for dense pedestrian detection. Finally, a lightweight shared convolutional semantic boundary detection head (LSCSBD) is designed. It utilizes shared convolutional layers across different feature hierarchies to significantly improve parameter efficiency and reduce redundant computation. Comparative experimental results on the CrowdHuman and WiderPerson datasets show that LDD-YOLO outperforms the comparative models in overall performance and achieves a balance between accuracy and efficiency.
    Reference | Related Articles | Metrics
    Abstract (344)
    PDF (201)
    Emotion Recognition from Physiological Signals: Review of Cross-Domain Transfer and Multimodal Fusion
    SHI Pengcheng, WANG Hailong, LIU Lin
    Journal of Frontiers of Computer Science and Technology    2026, 20 (5): 1261-1278.   DOI: 10.3778/j.issn.1673-9418.2505043
    Emotion recognition based on physiological signals has gradually emerged as a frontier research direction at the intersection of biomedical engineering and psychology in recent years. It not only provides objective, quantifiable new methods for the clinical diagnosis and efficacy assessment of affective disorders such as depression and anxiety, but also demonstrates broad application prospects in scenarios involving the perception of emotional states in healthy populations, including human-computer interaction, mental health monitoring, and intelligent driving. However, domain shift issues arising from individual physiological variations, inconsistent data acquisition equipment, and contextual diversity severely constrain models?? generalization capabilities in real-world cross-domain scenarios, forming the core bottleneck hindering the technology transition from controlled laboratories to practical applications. This paper focuses on two key technical approaches to enhance generalization: cross-domain transfer learning and multimodal fusion. Regarding cross-domain transfer, the paper critically examines domain shift challenges in electroencephalogram (EEG)-based emotion recognition, systematically summarizes the methodological evolution, applicable scenarios, and performance limitations of two primary approaches: domain adaptation and domain generalization. Concerning multimodal fusion, the paper categorizes three integration paradigms: feature-level, decision-level, and hybrid fusion, while elucidating the pivotal role of EEG-peripheral signal synergistic mechanisms in enhancing model accuracy and generalization. The paper concludes with an in-depth analysis of key challenges in current research and outlines prospective directions for future breakthroughs.
    Reference | Related Articles | Metrics
    Abstract (342)
    PDF (1298)
    Local Prototypes-Based Few-Shot Classification Model Embedded Hybrid Attention
    ZHOU Bin, XIAN Hao, QIN Yijia, ZHOU Hongbin
    Journal of Frontiers of Computer Science and Technology    2026, 20 (2): 584-595.   DOI: 10.3778/j.issn.1673-9418.2501016
    In few-shot learning tasks for novel class recognition under scarce data, prototype networks have gained significant attention. However, traditional methods relying on global averaging for prototype computation often suffer from background interference in extremely limited samples and insufficient coverage of local features. To address these issues, this paper proposes an enhanced prototype generation approach by integrating local feature representations and a hybrid attention mechanism. This paper extracts feature maps of samples through a backbone network and decomposes them into local descriptors. Element-wise mean aggregation is then applied to these descriptors, generating prototypes that fuse multi-scale details, thereby preserving global information while capturing rich contextual relationships and improving prototype discriminability and robustness. A hybrid channel-spatial attention mechanism is designed: the dual-pooling channel attention module employs global average pooling and max-pooling pathways to dynamically learn channel-wise weights for suppressing background noise; the enhanced spatial attention module evaluates regional relevance through a local feature similarity matrix, combined with channel statistics derived from global average and max-pooling, to collaboratively generate foreground-targeted spatial attention distributions. The two modules focus on critical local regions of images in complementary ways, enhancing prototype representativeness. Compared with existing methods, the approach demonstrates superior performance in classification tasks across varying granularities. For coarse-grained classification, accuracy improvements of 5.60 and 5.84 percentage points are achieved over baseline models under two experimental settings. In fine-grained tasks, the proposed method outperforms state-of-the-art methods by an average margin of 5.42 percentage points.
    Reference | Related Articles | Metrics
    Abstract (337)
    PDF (115)
    Review of Research on Multimodal Data Fusion Methods in Meteorology
    WU Ruofei, FANG Wei, JIANG Hongru, BAO Yansong
    Journal of Frontiers of Computer Science and Technology    2026, 20 (4): 905-922.   DOI: 10.3778/j.issn.1673-9418.2508007
    Advancements in multi-source observation technologies have ushered in the era of multimodal meteorological data. However, single-modal data is inherently limited in its ability to characterize the complex and dynamic atmospheric system, failing to meet the demand for more precise predictions. Consequently, integrating multimodal data to leverage complementary information and enhance performance has become a key research frontier in meteorology. This paper addresses the core challenge of effectively integrating meteorological data from different modalities by providing a systematic review of multimodal data fusion methods. First, this paper traces the evolution of multimodal fusion techniques, with a particular focus on deep learning-based strategies. It elaborates on the core concepts, architectural features, and advantages of mainstream models, such as encoder-decoder architectures, attention mechanisms, graph neural networks, and generative adversarial networks in the context of multimodal data fusion. Then, this paper comprehensively analyzes the current applications of this technology. Drawing on a systematic compilation of publicly available multimodal meteorological datasets, this paper reviews the technology??s application in key tasks, including precipitation nowcasting and predicting the paths and intensities of typhoons and tropical cyclones. It summarizes the research progress and effectiveness of different fusion methods in these scenarios. Furthermore, this paper analyzes the key challenges currently facing multimodal fusion in meteorology. Finally, based on the identified challenges, it outlines future research directions for the field.
    Reference | Related Articles | Metrics
    Abstract (332)
    PDF (2191)
    Differential Privacy in Federated Learning: Challenges and Prospects
    LIU Yi, JIANG Chengjie, YANG Songtao, ZHANG Lei, WU Shiwei
    Journal of Frontiers of Computer Science and Technology    2025, 19 (12): 3179-3201.   DOI: 10.3778/j.issn.1673-9418.2504088
    In federated learning, differential privacy serves as a key technology for addressing privacy concerns, yet it still faces multiple challenges in adapting to heterogeneous environments, personalized privacy design, and communication optimization. Focusing on the current research status of differential privacy in federated learning, this paper conducts an in-depth analysis around three typical application scenarios: differential privacy in heterogeneous environments, personalized differential privacy protection, and differential privacy-communication optimization mechanisms. Specifically, this paper explores the design ideas of differential privacy mechanisms under the conditions of statistical heterogeneity, device heterogeneity, and model heterogeneity respectively, and compares the technical paths and evolutionary trends of various methods in terms of perturbation strategies, budget allocation, and adaptation capabilities. On this basis, this paper further summarizes the performance of current differential privacy mechanisms in terms of computational efficiency, communication overhead, and model adaptability, concludes their advantages and disadvantages, and clarifies the applicable scenarios of different mechanisms through the comparison of typical algorithms. To more comprehensively evaluate the actual effects of these mechanisms, this paper also analyzes the impact of differential privacy strategies on model accuracy, communication rounds, and convergence speed in federated learning by combining experimental results, and reveals the trade-off relationship between privacy protection and utility improvement. Finally, based on the limitations of existing research, this paper prospects the joint optimization mechanisms of differential privacy in multi-dimensional heterogeneous and personalized scenarios.
    Reference | Related Articles | Metrics
    Abstract (324)
    PDF (2196)
    Research Review of Deep Learning in Skin Lesions Image Segmentation
    MENG Xiangfu, LI Jiaxun, YU Chunlin, LU Yunxuan
    Journal of Frontiers of Computer Science and Technology    2026, 20 (1): 21-39.   DOI: 10.3778/j.issn.1673-9418.2504019
    Skin lesions exhibit a wide variety of types and complex clinical manifestations, ranging from benign conditions to malignant melanomas. Early detection and accurate segmentation of these lesions are critical for the diagnosis and treatment of skin cancer, particularly in the early identification and localization of high-risk lesions such as malignant melanoma, which can significantly improve patient survival rates. In recent years, deep learning techniques have achieved remarkable progress in skin lesion image segmentation, greatly enhancing both accuracy and efficiency. This paper presents a comprehensive review of deep learning research in the field of skin lesion image segmentation. First, various skin lesion imaging modalities and commonly used public datasets are introduced, along with a summary of standard evaluation metrics. Then, addressing the prevalent issues of noise and artifacts in images, a detailed discussion on various image preprocessing and augmentation techniques is provided. Subsequently, deep learning-based segmentation methods are elaborated, covering U-Net, Transformer, SAM (segment anything model), Mamba, and multi-network fusion models. The main architectural designs, advantages, limitations, and segmentation performance of these models are comparatively analyzed. Finally, the current challenges and issues in this field are examined, and future research directions are proposed, aiming to provide valuable insights for the continued development of skin lesion image segmentation.
    Reference | Related Articles | Metrics
    Abstract (323)
    PDF (2375)
    Generative Diffusion Model for Incomplete Multimodal Emotion Recognition
    MA Fei, WANG Yuting, YANG Feixia, XU Guangxian
    Journal of Frontiers of Computer Science and Technology    2026, 20 (1): 206-216.   DOI: 10.3778/j.issn.1673-9418.2411057
    Human multimodal emotion recognition utilizes various heterogeneous modal data, such as language, vision, and sound, to perceive and understand human emotions. Compared with single modality, the complementary information in multimodal data helps to understand emotions more robustly. However, in practical multimodal scenarios, incomplete or missing modal information often exists, severely hindering the understanding of multimodal features and leading to a decline in emotion recognition accuracy. To address the issue of recognition accuracy decline caused by incomplete or missing modalities, which has not been effectively handled by previous multimodal emotion recognition methods, this paper proposes an incomplete multimodal emotion recognition method based on fusion generative diffusion models. This method aims to enhance emotion recognition accuracy by reconstructing incomplete modal data information. Firstly, a generative diffusion model based on cross-modal conditional stochastic differential equations is constructed. In the back-diffusion process, the available modal information is converted into dynamic constraints of the drift term through learnable projection to generate incomplete modal information features. Secondly, a two-way collaborative optimization framework of the incomplete modal generation network and the fusion reconstruction module is constructed. The joint objective function is used to realize the gradient backpropagation interaction of generation quality and feature fusion, and the emotional semantic consistency constraint of generated features and real features is established through a hierarchical attention mechanism. Results from several dataset tests indicate that the proposed multimodal emotion recognition method achieves superior performance in various incomplete modality scenarios.
    Reference | Related Articles | Metrics
    Abstract (320)
    PDF (95)
    Review of Deep Learning Based Lane Line Detection Methods
    HUANG Deqi, GUO Yanan, NI Zicong
    Journal of Frontiers of Computer Science and Technology    2026, 20 (6): 1545-1561.   DOI: 10.3778/j.issn.1673-9418.2508016
    With the rapid development of autonomous driving technology, lane detection, as the core task of environmental perception systems, faces key challenges such as occlusion, sudden changes in lighting, and geometric diversity in complex scenes. A systematic study is conducted on lane detection methods based on deep learning to address this issue, with a focus on analyzing the algorithm frameworks, implementation processes, and performance optimization strategies of 2D and 3D detection technologies. In terms of 2D detection, the limitations of traditional image processing methods are explored, and the applications of convolutional neural networks, fully convolutional networks, and Transformer architectures in feature extraction, instance segmentation, and lightweight design are elaborated in detail. In the aspect of 3D detection, the technical route based on LiDAR point cloud, stereo vision and multi-sensor fusion is described, and the advantages and disadvantages of different sensor combinations in geometric reconstruction and dynamic tracking are analyzed. By combining commonly used datasets and evaluation metrics, a standardized reference basis is provided for the standardized testing of algorithm performance. Analysis shows that multimodal data fusion and lightweight network design can effectively improve the stability of lane detection systems in complex scenarios such as occlusion and sudden changes in lighting, while 3D detection technology further enhances positioning accuracy by introducing spatial geometric information. Future research should focus on the construction of a universal modeling framework, innovation in multimodal complementary fusion mechanisms, enhancement of temporal reasoning capabilities in dynamic scenarios, and efficient and lightweight deployment of embedded platforms to promote the practical application of autonomous driving technology.
    Reference | Related Articles | Metrics
    Abstract (311)
    PDF (853)
    Fake News Detection Method Based on Attention-Guided Multimodal Feature Fusion
    DENG Xingyu, WANG Longye, ZENG Xiaoli, YE Hao, CHE Xihao
    Journal of Frontiers of Computer Science and Technology    2026, 20 (1): 194-205.   DOI: 10.3778/j.issn.1673-9418.2501008
    Existing multimodal fake news detection methods have limitations in utilizing image’s multilayered frequency-domain information and enabling deep interactions between modalities, making it difficult to fully exploit the latent features of images and the correlations between multimodal features, which in turn affects detection performance. Fake news images typically undergo multiple compression or tampering operations during their dissemination, which leads to abnormal frequency-domain responses. Traditional methods often rely on Fourier transforms to extract frequency-domain features. However, their global frequency analysis may miss local tampering traces and thus fail to achieve multi-scale feature decoupling. To better uncover and fully utilize these critical features and their inherent relationships, thereby improving fake news detection performance, an attention-guided multimodal feature fusion method for fake news detection (AGMFN) is proposed. This method models the multi-layered frequency-domain information of images using a wavelet-transform-based dual-path feature extraction module. It captures low-frequency global structures and high-frequency local anomalies through two-level wavelet decomposition, while enhancing detail features through feature-boosting convolution. Meanwhile, pre-trained models and frequency-domain feature extraction modules are employed to separately extract textual, visual, and frequency-domain features, constructing a joint framework for physical forensics and semantic clues. To enable multimodal feature fusion and capture deep correlations between different modalities, an attention mechanism-based long-sequence feature fusion module is designed, introducing an exponentially decaying weighting coefficient to model long-term dependencies between modalities and solve the temporal mismatch issue in traditional concatenation-based fusion. Through cross-modal attention, a hierarchical fusion of text, frequency-domain, and visual features is achieved, enhancing fake news detection capability while maintaining computational efficiency. Experimental results show that AGMFN achieves classification accuracies of 0.917 and 0.847 on the Weibo and Twitter datasets, respectively, outperforming existing baseline models. Visualization experiments further confirm that the fused multimodal features exhibit stronger generalization ability, thus improving the performance of fake news detection.
    Reference | Related Articles | Metrics
    Abstract (307)
    PDF (182)
    Survey of Word Embedding Models Research
    WEN Yongqi, YANG Ruopeng, TAO Yu, ZHONG Yihao, HUANG Bo
    Journal of Frontiers of Computer Science and Technology    2026, 20 (7): 1841-1860.   DOI: 10.3778/j.issn.1673-9418.2512024
    As a foundational technology in natural language processing, word embedding models map discrete linguistic symbols into continuous vector representations that computers can process. Their representational capacity and generalization performance directly impact the effectiveness of downstream tasks. Traditional word representation methods struggle to capture semantic relationships between words, while static word embedding methods fail to handle polysemy effectively. With the development of pre-trained language models and large language models, word embedding technology has gradually evolved from fixed vector representations to dynamic representations with contextual awareness. However, current research has yet to systematically incorporate the new paradigms of embedding models in the era of large language models. This paper comprehensively reviews the development trajectory of word embedding models, dividing their evolution into four stages based on typical technical paradigms: static word embeddings based on statistics, static word embeddings based on feedforward neural networks, dynamic word embeddings based on pre-trained language models, and dynamic word embeddings based on large language models. For each stage, the representative models, core principles, and their advantages and disadvantages are elaborated. The dynamic word embedding methods based on large language models are systematically categorized into three technical approaches: pooling, prompt engineering, and fine-tuning. The intrinsic and extrinsic evaluation methods for word embedding models and their interrelationships are examined. Addressing the current limitations of word embedding models, future research directions such as multimodal fusion, efficiency optimization, low-resource adaptation, and interpretability are proposed.
    Reference | Related Articles | Metrics
    Abstract (300)
    PDF (632)
    Advancements in Deep Learning-Based Vehicle Trajectory Prediction Research
    FANG Jinfeng, ZHANG Zhenwei, MENG Xiangfu
    Journal of Frontiers of Computer Science and Technology    2026, 20 (2): 346-366.   DOI: 10.3778/j.issn.1673-9418.2504029
    Vehicle trajectory prediction involves using artificial intelligence methods to forecast a vehicle??s future path and behavior over a given time period. In recent years, with the continuous growth in the number of vehicles, traffic-related issues have become more prevalent, making the ability to automatically perceive, understand, and predict the next route of a vehicle increasingly vital. Additionally, the widespread adoption of various traffic data collectors has led to the generation of vast amounts of vehicle trajectory data, making trajectory prediction highly valuable in several fields, including autonomous driving. This paper aims to provide a systematic review of vehicle trajectory prediction algorithms based on deep learning. First, it summarizes the key factors influencing prediction performance, such as dataset quality and driver intent. Then, traditional trajectory prediction approaches are reviewed. Building upon this foundation, this paper focuses on deep learning-based methods, including those based on recurrent neural networks, graph convolutional networks, graph attention networks, Transformers, and other deep generative models such as generative adversarial networks and variational auto-encoders. Subsequently, commonly used datasets and evaluation metrics in the field are introduced, and various deep learning methods are compared in terms of predictive performance and generalization capability. Finally, this paper discusses major challenges in vehicle trajectory prediction, such as environmental uncertainty and behavioral variability, and provides insights into potential future research directions.
    Reference | Related Articles | Metrics
    Abstract (295)
    PDF (2053)
    Review of Pedestrian Trajectory Prediction Based on Graph Neural Networks
    DU Ting, ZHUANG Xufei, WANG Yujie, LI Ziheng, LYU Jie, ZHI Yuanyuan, ZHAO Yupeng
    Journal of Frontiers of Computer Science and Technology    2026, 20 (4): 923-942.   DOI: 10.3778/j.issn.1673-9418.2507052
    Pedestrian trajectory prediction is a critical technology in scenarios such as autonomous driving and intelligent surveillance, where its accuracy directly impacts the decision-making quality of intelligent systems. Traditional methods have certain limitations in effectively modeling the complex social interactions and environmental factors influencing pedestrian movement. Graph neural networks (GNNs), with their strong relational modeling capabilities, have become an important tool in pedestrian trajectory prediction and have demonstrated significant advantages in various predictive tasks. This review systematically organizes GNN-based pedestrian trajectory prediction methods, categorizing them into three types based on model architecture: methods based on spatio-temporal graph neural networks, methods combining GNNs with generative models, and methods integrating GNNs with Transformers. The paper provides an in-depth analysis of the modeling mechanisms, advantages, and limitations of each approach, and summarizes the performance of several representative methods on public datasets. The challenges currently faced by the field are discussed, and potential future research directions are proposed.
    Reference | Related Articles | Metrics
    Abstract (295)
    PDF (1838)
    Review of Communication in Multi-agent Deep Reinforcement Learning from Seven Dimensions
    CHEN Rongmin, GUO Dabo, WU Hongkun, LI Chengxiang, HU Haixiao, LIU Lixiang
    Journal of Frontiers of Computer Science and Technology    2026, 20 (6): 1606-1626.   DOI: 10.3778/j.issn.1673-9418.2506028
    Deep reinforcement learning has achieved remarkable success in solving various sequential decision-making problems in single-agent machine learning. However, in real-world environments, tasks often involve multiple agents, leading to growing interest in multi-agent deep reinforcement learning (MADRL) techniques. MADRL faces challenges such as partial observability and environmental non-stationarity. Communication-based multi-agent deep reinforcement learning (CB-MADRL) addresses these issues by introducing communication mechanisms. Focusing on the communication module in CB-MADRL, and addressing the lack of systematic analysis of communication mechanisms in existing surveys, this paper classifies communication timing and proposes new classification methods for communication targets and message aggregation. A systematic seven-dimensional analytical framework is presented, encompassing communication timing, communication target, communication source, message aggregation, communication function, communication learning, and communication constraints. Based on this framework, recent CB-MADRL methods are categorized, a unified classification notation is designed, and the seven communication dimensions of each method are summarized. This paper systematically analyzes key challenges faced by the seven communication dimensions in real-world applications, including multi-modal and heterogeneous message processing, the “black-box” nature of communication protocols, and the coupling of real-world communication constraints, and proposes targeted research ideas and feasible approaches. Furthermore, it explores the interactions and synergistic optimization mechanisms among the communication dimensions, proposes a path for exploring optimal dimension combinations, and finally envisions future research directions.
    Reference | Related Articles | Metrics
    Abstract (288)
    PDF (683)
    Review of Fatigue Driving Detection Research
    ZHAO Ya, JIANG Liuyang, JIA Di, YAO Wenda, HUANG Shiwang
    Journal of Frontiers of Computer Science and Technology    2026, 20 (5): 1279-1293.   DOI: 10.3778/j.issn.1673-9418.2509031
    Fatigue driving is a major factor contributing to traffic accidents, making fatigue detection research crucial for ensuring road traffic safety. In recent years, advances in deep learning and intelligent vehicle technologies have improved fatigue detection techniques in terms of accuracy, real-time performance, and adaptability. This paper summarizes commonly used public fatigue driving datasets and their performance evaluation metrics. It categorizes fatigue detection methods by data acquisition approach and the evolution of discrimination dimensions, ranging from subjective to objective and from indirect to direct, covering physiological signals, observable behaviors, subjective evaluations, and expert assessments, and defines relevant discrimination criteria and calculation formulas. According to signal acquisition approach, fatigue detection methods are further divided into contact-based, non-contact-based, and hybrid approaches. Representative studies from the past five years using physiological signals, facial features, driver operational behaviors, and vehicle operational states are reviewed. A tabular summary outlines each method??s classification, contact type, applied datasets, and respective strengths and limitations. Finally, this paper analyzes the challenges of deploying fatigue detection systems in real-world scenarios, including computational constraints, insufficient cross-scenario generalization, and privacy compliance, providing valuable references for understanding the research trajectory and promoting practical applications.
    Reference | Related Articles | Metrics
    Abstract (285)
    PDF (1205)
    WMamba:Remote Sensing Weak Target Detection Based on Mamba and Edge-Aware Enhancement
    DONG Yan, WANG Zhen, GAO Guangshuai, GAO Junyu, LI Chunlei, LIU Zhoufeng
    Journal of Frontiers of Computer Science and Technology    2026, 20 (6): 1795-1808.   DOI: 10.3778/j.issn.1673-9418.2508062
    To address the challenges of detecting weak targets in remote sensing images, such as complex backgrounds, low-contrast blurry target boundaries, and low pixel proportions of small targets, a weak target detection algorithm named weak Mambal (WMamba), based on edge-aware enhancement, is proposed. A local Mamba module coupled with a local feature extraction mechanism is designed. This module integrates state space models with a local receptive field enhancement strategy, preserving the advantage of long-range dependency modeling while significantly improving the ability to capture detailed features. This structure is used to replace the C2f modules in both the backbone and neck networks, enabling the progressive extraction and fusion of multi-scale contextual information and local texture features. This process gradually builds feature representations with rich semantic hierarchies and strong discriminative power. To further enhance target contour information, an edge enhancement module is introduced. This module employs a strategy of differentiated positive and negative weight distribution to construct a gradient-level edge information extraction path. By amplifying high-frequency gradient responses and suppressing low-frequency background noise, it effectively highlights the contour features of blurry boundaries and weak targets, guiding the network to achieve salient representations in edge regions. To enable precise focus on key information within the enhanced features, a multi-dimensional attention module is designed. This module allocates attention weights simultaneously across three dimensions: spatial position, channel relationships, and convolutional kernels. This mechanism effectively suppresses interference from complex backgrounds and strengthens responses in target regions, thereby improving the distinguishability between weak targets and the background and enhancing detection accuracy. The proposed method achieves mean average precision (mAP) improvements of 5.5 and 3.2 percentage points compared with the baseline model YOLOv8n on the NWPU VHR-10 and UAVOD-10 datasets, respectively.
    Reference | Related Articles | Metrics
    Abstract (285)
    PDF (155)
    Research on Application Status and Progress of Large Language Model in Water Conservancy Field
    LI Xiaolong, XU Yi, MA Kai, CUI Wenchao, YI Shuang
    Journal of Frontiers of Computer Science and Technology    2026, 20 (8): 2223-2240.   DOI: 10.3778/j.issn.1673-9418.2512045
    Smart water conservancy is a crucial measure for promoting the efficient utilization of water resources and the modernization of water governance, and has become a core direction for the transformation and upgrading of the water conservancy sector. During the current construction process, it still faces problems such as heterogeneous multi-source data, insufficient coupling between mechanisms and data models, and limited level of business intelligence, which makes it difficult to fully support the demand for scientific decision-making in complex water conservancy scenarios. Large language model (LLM), relying on its advantages in knowledge representation, semantic reasoning and natural interaction, shows broad potential in empowering smart water conservancy. From the three core dimensions of technical system, business paradigm and system integration, this paper systematically reviews the research progress of LLM in the water conservancy field, sorts out the core technical systems of LLM empowering smart water conservancy, including knowledge enhancement, parameter optimization, interaction optimization, multi-modal fusion and agent development, and analyzes the dataset construction and evaluation methods of LLM in the water conservancy field. It deeply examines the transformation of business paradigms driven by LLM in core scenarios such as basin flood control, water resource allocation, and construction, operation and maintenance of water conservancy projects. It also summarizes three types of system integration models and technical pathways for business implementation, namely knowledge enhancement-driven, professional model coupling, and agent collaboration. Finally, the development trends of the in-depth integration of LLM and smart water conservancy are summarized and prospected, aiming to provide theoretical reference and practical experience for the intelligent and smart transformation of the water conservancy sector.
    Reference | Related Articles | Metrics
    Abstract (283)
    PDF (158)
    Research on Construction of Knowledge Graph for Tang-Bo Ancient Road Cultural Heritage Based on Large Language Models
    LIU Huarui, LUO Weiqun, ZHANG Zhaoji
    Journal of Frontiers of Computer Science and Technology    2026, 20 (3): 801-810.   DOI: 10.3778/j.issn.1673-9418.2510052
    To address the challenges of integrating multi-source heterogeneous data in the cultural heritage domain and the issues of traditional knowledge extraction methods relying on manual annotation and having a low degree of automation, this paper proposes a method for constructing cultural heritage knowledge graphs based on large language models (LLMs). Firstly, by integrating domain expert knowledge and cultural heritage industry standards, this paper constructs a knowledge ontology encompassing core cultural heritage concepts and their relationships. This ontology systematically defines spatio-temporal attributes, cultural attributes, and physical characteristics of cultural relics, as well as the associative relationships between them, forming a multi-dimensional, structured knowledge framework. Secondly, a joint extraction strategy with multi-stage task decomposition is designed. Leveraging the semantic understanding and generation capabilities of LLMs, the knowledge extraction process is decomposed into two subtasks: “entity-attribute extraction” and “triple extraction” guided step-by-step through prompt engineering. Simultaneously, a multi-dimensional verification mechanism is introduced to validate the generated triples from three aspects: logical consistency, domain normative compliance, and factual accuracy, ensuring the professionalism and reliability of the knowledge graph. Experiments show that the DeepSeek-R1 model achieves an F1-score of 86.25% in the extraction task, which is 3.13 percentage points higher than the BERT-BiLSTM-CRF model. Ablation studies also verify the effectiveness of each module. This method provides efficient and automated technical support for the digital preservation of cultural heritage and interdisciplinary research.
    Reference | Related Articles | Metrics
    Abstract (281)
    PDF (122)
    Dynamic Emoji Perception and Bidirectional Cross-Modal Attention for Large Language Model Sentiment Analysis
    LIU Bowen, LI Fengqi, LI Shenghui, LIANG Shu'ao, WANG Deguang
    Journal of Frontiers of Computer Science and Technology    2026, 20 (7): 2064-2078.   DOI: 10.3778/j.issn.1673-9418.2510010
    Social media platforms (such as Twitter) serve as vital conduits for user sentiment and public discourse in modern society. Their vast volumes of user-generated content carry rich personal emotions and collective opinions. The real-time nature, short-text characteristics, diversity, and unstructured nature of this data provide a rich foundation for sentiment analysis research while also presenting significant technical challenges. Traditional sentiment analysis methods based on a single text modality struggle to fully capture the complex emotional information embedded in social media. They particularly overlook the significant emotional characteristics conveyed by visual modalities (e.g., images) and symbolic modalities (e.g., Emojis), leading to significant bottlenecks in model performance. Addressing the issues of semantic dynamism and text-image heterogeneity in social media, this paper focuses on the Twitter platform and proposes an emotion analysis framework integrating large language models with dynamic Emoji perception and bidirectional cross-modal attention. This paper designs a bidirectional cross-modal attention mechanism incorporating dynamic modality quality assessment. This mechanism adaptively adjusts fusion weights based on the informational quality and reliability of text and images, outperforming simple feature concatenation or fixed-weight attention fusion. This paper also proposes a context-aware dynamic Emoji semantic parsing and text augmentation method driven by large language models. This approach leverages the context-aware capabilities of large language models to overcome the static semantic limitations of Emoji symbols, enhance textual sentiment features, and effectively handle ironic contexts. Utilizing Llama3.1-8B as the base model, combined with advanced low-rank adaptive fine-tuning techniques and hyperparameter optimization strategies, the method achieves high accuracy in sentiment recognition while significantly reducing computational resource consumption. Experimental design involves ablation studies validating the effectiveness of multimodal fusion and dynamic Emoji processing. Comprehensive evaluations are conducted on the SemEval-2017 and Twitter-2015/2017 datasets. Experimental results demonstrate improvements across all evaluation metrics for the fine-tuned model. Particularly for samples containing sarcastic content, the model achieves notably higher recognition accuracy due to the complementary nature of multimodal information.
    Reference | Related Articles | Metrics
    Abstract (279)
    PDF (118)
    YOLO-Mamba Contraband Detection Method Integrating Global Perception and Multi-scale Collaboration
    SHENG Chunlei, LIU Chengkai, LI Zelong, LU Shuhua
    Journal of Frontiers of Computer Science and Technology    2026, 20 (4): 1169-1180.   DOI: 10.3778/j.issn.1673-9418.2503063
    To address the challenges including spatial multi-scale variations, target occlusion, and the high false negative and false positive rates of small contraband objects in X-ray security images, this paper proposes a YOLO-Mamba contraband detection algorithm that integrates global perception and multi-scale collaboration. Using YOLOv10 as the baseline, a gated structure-aware selective block (GSSBlock)-enhanced state-space model is incorporated into the backbone to optimize spatial information modeling and effectively focus on key region features. Additionally, a multi-residual connected pooling (M-RCP) structure is designed to enhance the perception of both global and edge information, improving foreground-background differentiation and mitigating the impact of complex background interference. In the neck, a deep feature fusion pyramid network (DFFPN) is introduced, employing bidirectional cross-scale interactions and multi-level feature fusion to strengthen multi-scale feature perception and reduce false positives and false negatives. A dual-branch depthwise separable convolution (DWConv) fusion module is utilized to extract information from different receptive fields, effectively capturing fine-grained details of small contraband objects while maintaining computational efficiency. The proposed method is trained and evaluated on three public datasets: OPIXray, HIXray, and SIXray, achieving mAP50 scores of 94.0%, 82.9%, and 94.6%, with improvements of 5.7, 2.7, and 4.6 percentage points over the baseline. Experimental results demonstrate that the proposed approach outperforms various state-of-the-art algorithms while maintaining a lower complexity, effectively balancing detection accuracy and computational efficiency, making it a competitive solution for contraband detection in X-ray security screening.
    Reference | Related Articles | Metrics
    Abstract (274)
    PDF (140)
    Research on Education Text Classification Based on Prompts Learning and Large Language Models
    RONG Lu, YU Mei, ZHAO Dongming, WANG Yonggang, ZHANG Yazhou
    Journal of Frontiers of Computer Science and Technology    2026, 20 (4): 1006-1018.   DOI: 10.3778/j.issn.1673-9418.2508002
    Educational text classification is a core task in intelligent education, with wide applications in textbook content analysis, automatic exam-question categorization, and educational-resource recommendation. Compared with general texts such as news or reviews, educational texts are characterized by dense terminology and rich structured knowledge, which substantially increase the difficulty of semantic understanding and transfer learning. Under low-resource and small-sample conditions, mainstream approaches that rely on large-scale annotated data and supervised fine-tuning suffer from severe data scarcity. This paper proposes a text-classification method tailored for low-resource and small-sample scenarios, based on a prompt-tuning strategy that facilitates the transfer of prior knowledge from pre-trained language models (PLMs) and large language models (LLMs). This paper designs three types of prompt templates: discrete, continuous, and hybrid. It incorporates a bidirectional long short-term memory network (BiLSTM) to model dependencies among prompt tokens, thereby improving the model’s task-specific semantic adaptation. Experiments on four representative benchmarks (EduBooks, AGNews, MELD, and IMDB) show that the proposed method outperforms eight strong baselines on F1 by 3.7%, 3.3%, 5.7%, and 1.1%, respectively. The results indicate that the proposed approach can efficiently and accurately classify diverse text types with strong generalizability, without fine-tuning the parameters of the pre-trained models. Ablation studies further validate the effectiveness of the parameter-freezing strategy.
    Reference | Related Articles | Metrics
    Abstract (273)
    PDF (132)
    Recent Advances in Speech-Driven Gesture Generation
    ZHANG Yayu, WEN Yuhui, ZHANG Xinyu, JING Liping
    Journal of Frontiers of Computer Science and Technology    2026, 20 (3): 611-624.   DOI: 10.3778/j.issn.1673-9418.2505081
    In interpersonal communication, gestures enrich verbal information and facilitate information delivery. Speech-driven gesture generation aims to automatically synthesize natural, realistic, and contextually appropriate sequences of gestures conditioned on speech input. This research direction has attracted widespread attention in fields such as computer graphics and computer vision, holding significant application value in domains including film animation production, human-computer interaction, and virtual reality. Early rule-based methods suffer from inefficiency, while regression methods, despite improving generation efficiency, often result in gestures with repetitive motion patterns and limited expressiveness.?In recent years, generative models have further advanced this field, effectively enhancing the quality and diversity of generated gestures. Regarding speech-driven gesture generation methods based on generative models, this work summarizes and categorizes relevant research on generative adversarial networks, variational autoencoders, and diffusion models, analyzing their respective applications, advantages, and disadvantages in gesture generation. It further explores the controllability of speech-driven gesture generation in emotion expression, semantic consistency, and style transfer. Moreover, collaborative generation research combining facial expressions and gestures is discussed. Additionally, commonly used datasets and evaluation metrics are introduced, followed by experimental comparative analysis of representative methods. Finally, this paper concludes by summarizing the challenges in the field of speech-driven gesture generation and outlining future research trends.
    Reference | Related Articles | Metrics
    Abstract (271)
    PDF (2053)