Most Download articles

Published in last 1 year| In last 2 years| In last 3 years| All| Most Downloaded in Recent Month | Most Downloaded in Recent Year|

Published in last 1 year

Please wait a minute...


For Selected: Download Citations EndNote Ris BibTeX Toggle Thumbnails

Select

Improved YOLOv13 for Infrared Remote Sensing Small Object Detection LI Ping, CHEN Jifeng Computer Engineering and Applications    2026, 62 (7): 131-142.   DOI: 10.3778/j.issn.1002-8331.2509-0101 Abstract （522）      PDF（pc） （5813KB）（4337）       Knowledge map    Save To address the challenges of small object size, low signal-to-noise ratio, and susceptibility to complex background interference in infrared remote sensing, the YOLOv13 model is improved to meet the requirements of real-time and lightweight detection. A multi-level feature aggregation module (MFAM) is constructed to aggregate hierarchical information from different semantic depths and spatial resolutions in a bottom-up manner, while adaptively recalibrating their contributions to alleviate the dilution of small objects in deep semantic layers. A dual-path fusion pyramid network (DFPN) is designed, where a top-down semantic enhancement path and a bottom-up detail refinement path interact in a complementary manner to achieve cross-scale information circulation, thereby enhancing the separability of weak thermal targets. The proposed context-aware fusion block (CAFBlock) adopts a dual-branch structure of global self-attention and local depthwise convolution to jointly model long-range dependencies and fine-grained local features. In addition, it integrates a dual-path processing strategy of dilated convolution with multiple receptive fields and depthwise convolution for local details, combined with a gated fusion mechanism, to comprehensively strengthen multi-scale context modeling. Comparative experiments on the SIRST and HIT-UAV datasets demonstrate that the improved model achieves 90.06% and 64.37% AP, with relative gains of 7.65 percentage points and 8.55 percentage points, respectively, which verifies the effectiveness and feasibility of the proposed approach. Reference | Related Articles | Metrics

Select

Research Review of UAV Recognition Based on Multi-Modal Fusion LI Minshu, ZHOU Mohan, ZHI Ruicong Computer Engineering and Applications    2025, 61 (21): 1-14.   DOI: 10.3778/j.issn.1002-8331.2501-0014 Abstract （1188）      PDF（pc） （17525KB）（983）       Knowledge map    Save With the rapid development of UAV technology, the application of related technologies is increasing, but it also brings many security risks and regulatory problems. As an important means to deal with these challenges, Anti-UAV detection technology has received extensive attention. Traditional UAV detection methods rely on a single modal data, such as vision, audio, radar and radio frequency signals, etc. However, these single modal data can obtain limited information in complex scenes. In recent years, deep learning methods have made good progress in the field of small object detection, and the related research of multi-modal fusion technology has further improved the accuracy and robustness of object detection. This paper reviews the research progress in the field of UAV detection, and focuses on summarizing the current research status of multimodal fusion technology. In addition, the evaluation indicators and public datasets of related UAV detection are sorted out, the limitations of existing technologies are analyzed, and the research directions for improving detection accuracy and robustness in the future are also pointed out. Reference | Related Articles | Metrics

Select

Survey of Medical Image Registration Based on Unsupervised Learning SHEN Longfeng, YU Yexun Computer Engineering and Applications    2026, 62 (18): 18-32.   DOI: 10.3778/j.issn.1002-8331.2509-0271 Abstract （884）      PDF（pc） （3330KB）（640）       Knowledge map    Save Medical image registration is a fundamental task in the field of medical image analysis and has been widely used in clinical applications such as disease progression tracking, atlas construction, statistical analysis, and surgical navigation. However, traditional registration algorithms are limited by their relatively low computational efficiency, making it difficult to meet the diverse application demands nowadays. In recent years, deep learning-based medical image registration methods have gradually become a major research focus in the industry. Furthermore, due to the difficulty in obtaining annotated datasets with realistic deformation fields, how to construct registration models through unsupervised learning has emerged as a prominent research topic. By reviewing relevant literature on unsupervised medical image registration over the past decade, this paper categorizes the mainstream registration frameworks into three types：single-stream, cascade, and pyramid architectures. Based on this categorization, it systematically elaborates on the development trajectory, key deep learning technologies, and future directions of unsupervised medical image registration. Reference | Related Articles | Metrics

Select

Survey on Multi-Label Learning Under Complex Data Environment DU Guodong, WANG Ao, LU Pengwei, YE Qianzhi, ZHANG Jia Computer Engineering and Applications    2026, 62 (11): 41-61.   DOI: 10.3778/j.issn.1002-8331.2507-0192 Abstract （503）      PDF（pc） （1120KB）（564）       Knowledge map    Save Multi-label learning is an important learning paradigm in the field of machine learning, which shows advanced performance in many applications. Data with the main characteristics of large-scale, incomplete information, multiple dimensions and dynamic evolution are bringing many challenges to multi-label learning. It is necessary to carry out a more refined design of the multi-label learning method according to the characteristics of different learning tasks under complex data environment. Therefore, this paper analyzes the source of the complexity of multi-label data and reviews the research progress of multi-label learning under complex data environment. Firstly, the paper introduces the problem definition, learning form and development trend of multi-label learning in detail. Next, the complexity of multi-label data from three levels：label, feature and sample is discussed. Then, the summarized, elaborated and analyzed multi-label learning methods under complex data environment are provided. Finally, the challenges and the possible future application directions of multi-label learning are summarized. Reference | Related Articles | Metrics

Select

YOLO-PD:Lightweight Real-Time Pedestrian Detection Algorithm CHEN Shengbao, SHI Longzhao Computer Engineering and Applications    2026, 62 (6): 122-133.   DOI: 10.3778/j.issn.1002-8331.2507-0218 Abstract （408）      PDF（pc） （13743KB）（544）       Knowledge map    Save To address the limited resources of edge devices and the low detection accuracy of pedestrian detection algorithms on small objects, multi-scale objects, and geometrically deformed objects caused by overlapping occlusions, YOLO-PD is proposed which is a YOLOv8-based pedestrian detection algorithm suitable for edge deployment. For small objects, a partial Transformer block (PTB) is designed to enhance feature extraction. Hybrid structure of PTB maintains efficient feature extraction while reducing computational costs. For multi-scale objects, a pyramid shared dilation convolution (PSDC) is introduced, leveraging shared-weight multi-scale dilation convolutions to improve multi-scale feature extraction while minimizing module size. For geometrically deformed objects, a light deformable dynamic head (LDDH) is developed. It dynamically adjusts weighting factors to improve detection accuracy and employs deformable convolutions to better capture features of deformed objects. Experimental results show that compared to the baseline model YOLOv8n, YOLO-PD achieves 2.9 and 1.9 percentage points higher mAP50 on the custom pedestrian detection datasets COCO-Person and VOC-Person, respectively, while reducing parameters by 34.3%. On the public dataset WidePerson, it improves mAP50 by 1.8 percentage points and mAP50:95 by 1.1 percentage points. The algorithm excels in pedestrian detection with high accuracy, strong generalization, and minimal parameters, making it ideal for edge devices. Reference | Related Articles | Metrics

Select

Enhanced YOLOv11s for Detecting Small Objects in Aerial Images GUO Jialin, CAO Yunfeng Computer Engineering and Applications    2026, 62 (8): 80-92.   DOI: 10.3778/j.issn.1002-8331.2507-0111 Abstract （389）      PDF（pc） （7612KB）（468）       Knowledge map    Save Aiming at the challenges of high-dynamic scenarios, limited target feature information, and constrained onboard computational resources in small object detection from UAV aerial images, this paper proposes a detection method for small aerial objects based on an improved YOLOv11s. Firstly, a dense cross-path feature pyramid network (DCP-FPN) is introduced as the Neck structure, where the large-object detection layer is removed and a small-object detection layer is added. By adopting dense connections to maximize feature utilization efficiency, the model significantly reduces the number of parameters while effectively enhancing the capability to capture small object features. Secondly, a novel lightweight convolutional module named Ghost-DSConv is proposed, which integrates dynamic channel adjustment and dual-branch feature interaction. The main branch employs depthwise separable convolution (DSConv), supplemented by a low-cost branch composed of group convolutions to generate supplementary features. This design extracts spatial features at a lower computational cost and enhances feature diversity. Finally, the paper proposes a lightweight small-target attention mechanism module (lightweight small-target attention, LSTA), which introduces a triple attention reinforcement mechanism to significantly enhance the model??s detection performance for small targets with minimal computational overhead. The experimental results on the VisDrone2019-DET dataset demonstrate that the improved algorithm proposed in this paper achieves 44.8% mAP50 and 27.4% mAP50-95. The enhanced YOLOv11s model has only 3.51×106 parameters and is deployed on an RK3588s?based embedded platform, where it successfully meets real-time detection requirements, achieving a favorable balance between accuracy and efficiency. Additionally, the proposed algorithm??s strong generalization capability and effectiveness are further validated on the CARPK and Tinyperson datasets. Reference | Related Articles | Metrics

Select

Review on Development of Spatio-Temporal Graph Neural Networks for Traffic Flow Prediction YAN Jiahe, LI Honghui, SUN Jing, LIU Jie, ZHANG Junwen, YANG Xiaorui, XU Yi Computer Engineering and Applications    2025, 61 (22): 1-19.   DOI: 10.3778/j.issn.1002-8331.2502-0225 Abstract （897）      PDF（pc） （1418KB）（437）       Knowledge map    Save In recent years, the application of deep learning in traffic flow prediction has attracted wide attention, especially the spatio-temporal graph neural network has achieved remarkable success in capturing spatio-temporal dependencies and predicting traffic characteristics. Although some reviews have examined the application of spatio-temporal graph neural networks, most of these studies focus primarily on application scenarios and fail to provide an in-depth analysis from the perspective of model design. Furthermore, a unified model classification framework is absent. This paper proposes a hierarchical classification method that considers the key elements such as module selection, fusion mechanism, architecture design, and training strategy. The spatio-temporal graph neural networks can be divided into six categories, namely recurrent graph convolutional network, spatio-temporal fully convolutional network, spatio-temporal attention network, spatio-temporal encoder network, spatio-temporal hybrid architecture network, and spatio-temporal networks with additional strategies. For each category, the unique model construction and fusion mechanisms are analyzed in detail, and the different model variants are compared. By analyzing both representative and recent works, the development trend of spatio-temporal graph neural networks is discussed, and the code addresses of open-source models are provided. Subsequently, the commonly used public datasets are gathered, and the performance of the latest advanced models is visually analyzed by comparing the results of previous experiments. Finally, the development opportunities and challenges in this field are summarized to offer insights for future research. Reference | Related Articles | Metrics

Select

Review of Monocular Vision Object Detection Based on Deep Learning LIU Guichao, WANG Huaiguang, REN Guoquan, WU Dinghai Computer Engineering and Applications    2026, 62 (1): 1-19.   DOI: 10.3778/j.issn.1002-8331.2504-0330 Abstract （433）      PDF（pc） （2212KB）（426）       Knowledge map    Save Monocular visual object detection, characterized by its low hardware cost and high real-time performance, has gradually become a core technology in fields such as autonomous driving and intelligent surveillance, playing an indispensable role. However, issues such as geometric ambiguity, occlusion robustness, and small object detection accuracy remain significant bottlenecks in current research. This paper primarily focuses on algorithm-level advancements, systematically quantifying the progress in monocular visual object detection technology from three perspectives: algorithm evolution, performance evaluation, and lightweight design. Firstly, single-stage detection algorithms are ?deconstructed into classic convolutional architectures and Transformer architectures, analyzing and summarizing their structural innovations and performance bottlenecks, revealing the trade-off patterns between accuracy, speed, and complexity. Secondly, from the three levels of network design, algorithm optimization, and model compression, this paper explores the integration strategies of lightweight technologies with object detection algorithms. It also consolidates multi-dimensional evaluation metrics from the three main official datasets used for training and evaluation in object detection, building a cross-model comparison framework based on the MS-COCO-2017 dataset to conduct a horizontal performance comparison of different single-stage detectors. Finally, the paper looks ahead to cutting-edge directions such as multimodal fusion and lightweight improvements, aiming to provide systematic references for the engineering application and theoretical breakthroughs of monocular visual object detection algorithms. Reference | Related Articles | Metrics

Select

UAV Small Target Detection Method Based on Multi-Modal Fusion Network YAO Jilin, LIU Hongzhe, ZHANG Cheng, LU Lu Computer Engineering and Applications    2026, 62 (4): 273-283.   DOI: 10.3778/j.issn.1002-8331.2508-0255 Abstract （267）      PDF（pc） （42031KB）（421）       Knowledge map    Save To address the problem of low accuracy and high false detection rate in UAV small target detection under low-light and complex background conditions，where single-modality detection methods often fail to achieve satisfactory performance, this paper proposes a small target detection algorithm based on a dynamic interaction fusion strategy, named YOLOv8-DF. A dual-stream feature extraction network is designed, which extracts infrared and visible features in parallel. The receptive field attention convolution (RFAConv) is introduced into the backbone to enhance the multi-scale perception capability of the model. To further capture global contextual information, a remote information enhancement module (RIEM) based on attention mechanism is proposed. A dynamic feature interaction module (DFIM) is designed to dynamically adjust the weights of advantageous information across modalities, thereby achieving deep feature fusion. Experimental results show that, on the DroneVehicle dataset, YOLOv8-DF improves mAP50 by 12.3 and 13.4 percentage points compared with the single-modality baseline models YOLOv8_RGB and YOLOv8_IR, respectively. Compared with the PSFusion detection method, mAP50 is improved by 8.4 percentage points. Furthermore, computational efficiency and generalization experiments conducted on the LLVIP public dataset demonstrate that the proposed method achieves both competitive efficiency and strong generalization performance. Reference | Related Articles | Metrics

Select

Review of Security Protection of Industrial Control Systems Based on Digital Twin WANG Kunkun, LI Yahui, DU Xin, XU Peihang, ZHOU Chunjie Computer Engineering and Applications    2026, 62 (6): 1-15.   DOI: 10.3778/j.issn.1002-8331.2508-0252 Abstract （383）      PDF（pc） （1788KB）（391）       Knowledge map    Save Digital twin technology constructs high-fidelity dynamic models of physical entities in a digital space, enabling real-time simulation of their behaviors in virtual spaces, and facilitating interactive coevolution between the digital twin and its physical counterpart. It offers a novel perspective for addressing security challenges in industrial control systems (ICSs). This paper outlines the multiple challenges faced by ICSs in ensuring system availability. The fundamental principles, evolutionary history, and key enabling technologies of digital twins are delineated. The applications of digital twin technology in ICSs security are surveyed, emphasizing progress in system emulation, protective architectures, anomaly detection, and risk assessment, alongside an assessment of current research shortcomings. The future development directions of digital twins in the information security protection of ICSs are summarized and prospected. Reference | Related Articles | Metrics

Select

Harmony-YOLO11：Lightweight Small Object Detector via High-Frequency Enhancement and Feature Guidance SUN Zhongyi, WANG Dong, CAO Guogang, ZONG Ming Computer Engineering and Applications    2026, 62 (10): 111-122.   DOI: 10.3778/j.issn.1002-8331.2509-0361 Abstract （324）      PDF（pc） （2129KB）（347）       Knowledge map    Save Small object detection in drone-captured images faces challenges including scale imbalance, limited visual information, and constrained computational resources. Existing detection models often struggle to balance accuracy and complexity for this task. To address these issues, this paper proposes Harmony-YOLO11, an improved YOLO11n-based algorithm for drone-captured small object detection. It introduces a high-frequency enhancement and cross-scale adaptive module (HCM) to strengthen small object edges and focus on object regions, enhancing model adaptability. An efficient feature guide feature pyramid network (FG-FPN) is developed with a feature fusion guide block (FFGB), achieving lightweight and effective feature fusion through simplified network paths. The C3K2_CCA module integrates CoordConv and CoordATT to improve spatial perception for small objects. A lightweight downsampling module MGC (maxpooling-ghost convolution) further reduces model complexity. Experiments on the VisDrone2019 dataset demonstrate that Harmony-YOLO11 improves mAP50 and mAP50-95 by 5.10 and 3.12 percentage points respectively compared to YOLO11n, while reducing parameters and model size by 31% and 25%. Additional evaluations on the TinyPerson dataset confirm the generalization and robustness of the algorithm. Reference | Related Articles | Metrics

Select

ACED-YOLO：Multi-Feature Enhancement and Scale-Adaptive Intelligent Driving Detection Algorithm XUE Zhiyang, PAN Lihu, SONG Zeju, ZHANG Rui, ZHANG Guoyou Computer Engineering and Applications    2026, 62 (10): 230-244.   DOI: 10.3778/j.issn.1002-8331.2507-0369 Abstract （316）      PDF（pc） （5734KB）（327）       Knowledge map    Save Object detection is crucial for the accurate identification and localization of various traffic objects, such as motor vehicles, non-motorized vehicles, and pedestrians, in complex traffic scenarios for intelligent driving systems, which directly affects driving safety and reliability. However, existing methods still have limitations in edge feature representation, multi-scale feature processing, and long-range dependency capturing, which affect detection performance and computational efficiency. To address these issues, a lightweight intelligent driving detection algorithm ACED-YOLO is proposed with multi-dimensional feature enhancement and scale adaptability. A bidirectional axial context-aware feature integration neck network (BC-FIN) is developed. This network captures global context synchronously in both horizontal and vertical dimensions and explicitly characterizes rectangular key regions using axial features. A spatial feature reconstruction and dynamic interpolation fusion design is proposed to obtain rich multi-level features and contextual information. An edge Sobel enhanced spatial integration module (E-SESIM) is constructed, optimizing the C3k2 module. By paralleling Sobel and convolution branches, edge and spatial features are effectively fused, and a residual structure is embedded to suppress gradient vanishing, enhancing the object contour representation capability. An adaptive dynamic multi-scale detection head (DMDH-DCNv4) is designed, utilizing a multi-level attention mechanism and deformable convolutions to dynamically adjust feature extraction. This design adapts to multi-scale and multi-dimensional spatial features, improving both detection speed and feature capture efficiency. Experiments on the KITTI dataset show that ACED-YOLO achieves an mAP@0.5 of 85.2%, which is a nearly 10 percentage points improvement over YOLOv11n. Additionally, ACED-YOLO demonstrates excellent performance in detection accuracy and computational efficiency on the TinyPerson and UrbanNMV custom datasets. This method shows a significant advantage in object detection tasks for intelligent driving scenarios. Reference | Related Articles | Metrics

Select

Review of Question Answering Techniques for Knowledge Graph QIAN Shenyi, FU Bowen, LI Daiyi, LIANG Yaoyao Computer Engineering and Applications    2025, 61 (23): 1-23.   DOI: 10.3778/j.issn.1002-8331.2501-0066 Abstract （392）      PDF（pc） （1714KB）（326）       Knowledge map    Save Intelligent question answering is a key technology to obtain demand information accurately and quickly from massive data. In recent years, intelligent question answering technology has achieved remarkable development, such as problem-based information extraction technology, semantic understanding technology and vector modeling method. However, with the rapid development of intelligent question answering technology, people are eager to have a reasonable division of intelligent question answering model to facilitate the use of users in different fields. In order to divide intelligent question answering model reasonably, it is convenient for researchers in the field of intelligent question answering to conduct in-depth research. Through the investigation of relevant literature in the field of knowledge graph Q&A, this paper summarizes the key technologies of knowledge graph Q&A, including entity linking and knowledge embedding, and introduces the related concepts and processing flow of knowledge graph Q&A in detail. In addition, according to different methods, knowledge graph-oriented question answering techniques are divided into three main categories: semantic parsing, information retrieval and large language model-based methods, and their advantages and disadvantages are introduced and the evaluation indexes of knowledge graph question answering models are summarized respectively. Finally, some suggestions and thoughts are put forward for the existing problems and the future development direction of knowledge graph question-answering technology. Reference | Related Articles | Metrics

Select

Review on Research and Applications of Deep Learning in Music Generation ZHAO Mingyuan, ZHOU Ruohua, YUAN Qingsheng, ZHOU Zhitong Computer Engineering and Applications    2026, 62 (10): 1-25.   DOI: 10.3778/j.issn.1002-8331.2505-0135 Abstract （302）      PDF（pc） （2110KB）（322）       Knowledge map    Save Music generation is an important interdisciplinary research direction in artificial intelligence and computer science, which has achieved remarkable progress in recent years with the breakthroughs of deep learning technologies. This paper presents a systematic review of research and applications of deep learning in the field of music generation. Music representations are categorized into three types：symbolic representation, audio representation, and multimodal representation. The architectural characteristics and performance comparisons of mainstream deep generative models, including LSTM, VAE, GAN, Transformer, and diffusion models, are outlined. Music generation tasks are divided into three categories according to generation requirements：symbolic music generation, audio music generation, and vocal music generation, with representative methods and key technological advancements in each task category systematically sorted out. On this basis, commonly used datasets suitable for different generation tasks, as well as subjective and objective evaluation metric systems, are summarized, and the performance of representative models is compared and analyzed through experimental data. The main challenges currently faced in music generation are discussed, including difficulties in music information modeling, data scarcity, lack of unified evaluation systems, and ethical and commercialization issues. Future application prospects in fields such as music education and psychotherapy are also explored. Reference | Related Articles | Metrics

Select

Review of Research on Multimodal Fusion Event-Based Camera SLAM Technology ZHAO Junyang, YU Han, ZHANG Zhili, ZHU Huixin, LYU Shenhua, ZHANG Kefan Computer Engineering and Applications    2026, 62 (9): 1-19.   DOI: 10.3778/j.issn.1002-8331.2505-0237 Abstract （353）      PDF（pc） （3751KB）（311）       Knowledge map    Save Event cameras are an emerging type of sensor that offer high dynamic range, low latency, high temporal resolution, and low data redundancy. Applying event cameras to simultaneous localization and mapping (SLAM) technology can fully leverage their advantages, breaking the limitations of traditional sensors in various application scenarios. Therefore, this paper reviews the current research status of multimodal fusion event camera SLAM technologies and analyzes the characteristics and limitations of different approaches. Firstly, the paper introduces the fundamental principles of SLAM technology and event cameras, elaborating on the advantages of event cameras over traditional vision sensors. Secondly, from the perspectives of unimodal event camera SLAM (monocular, stereo) and multimodal fusion event camera SLAM (event-inertial, visual-inertial, multi-source sensor fusion), the paper analyzes the strengths and weaknesses of both in terms of real-time performance, application scenarios, and computational requirements, and summarizes their developmental trajectories. Subsequently, it discusses the impact of unimodal and multimodal fusion event cameras on system performance and provides a comparative analysis of the performance of some multimodal fusion event camera systems. Finally, this paper outlines future research directions for multimodal fusion event camera SLAM technology, such as optimization of multimodal event SLAM algorithms, enhancement of generalization capabilities, and expansion of multi-source heterogeneous sensor fusion architectures, aiming to provide references and inspiration for the application of event cameras in the field of SLAM. Reference | Related Articles | Metrics

Select

Survey on Traffic Engineering in Software-Defined Networking Based on Reinforcement Learning LIU Yanfei, WANG Chengjin, LI Chao Computer Engineering and Applications    2025, 61 (24): 1-28.   DOI: 10.3778/j.issn.1002-8331.2412-0248 Abstract （389）      PDF（pc） （1399KB）（306）       Knowledge map    Save Software-defined networking（SDN）, with its global and centralized management architecture, has brought revolutionary changes to the management of complex and dynamic networks, and has also created favorable conditions for network traffic engineering. Concurrently, reinforcement learning has garnered significant attention due to its pronounced advantages in decision optimization. The integration of reinforcement learning with the unique architecture of SDN and its application to SDN traffic engineering holds substantial practical significance. Firstly, from both theoretical and practical perspectives, based on the trajectory of technological development, the paper reviews the advancements in reinforcement learning, deep reinforcement learning, and multi-agent deep reinforcement learning in SDN traffic engineering. Additionally, it conducts a thorough synthesis and analysis of existing research outcomes across various dimensions, including methodological categorization, network scenarios, reinforcement learning algorithms, and traffic engineering objectives, providing a multidimensional perspective on the integration of reinforcement learning with SDN traffic engineering. Subsequently, it further summarizes the research progress of reinforcement learning combined with other technologies, demonstrating its potential to enhance the performance of traffic engineering. Ultimately, based on a summary of the current research progress, the paper analyzes the challenges faced and proposes future research directions, providing some reference for deepening exploration in this domain. Reference | Related Articles | Metrics

Select

Survey on Question Answering Systems and Evaluation in the Era of Large Models CUI Longfei, WANG Zongshui, BAO Yingxu, ZHAO Hong Computer Engineering and Applications    2026, 62 (5): 1-17.   DOI: 10.3778/j.issn.1002-8331.2503-0004 Abstract （304）      PDF（pc） （1660KB）（299）       Knowledge map    Save In the era of large models (LMs), question answering (QA) systems exhibit new characteristics. This paper reviews QA system features and evaluation methods. From the stages of training data for question-answering model inference training, pre-training frameworks, model post-processing, and efficient model fine-tuning, it contrasts the early “pursuit of data and parameter scale” training methods with the current “emphasis on data and model efficiency, ” and systematically analyzes the new characteristics of large-model-based automatic question-answering systems. The paper summarizes existing QA evaluation methods, reviewing HELM (holistic evaluation of language model) regarding datasets, metrics, and quantitative assessment for QA tasks. Future research on LM-based QA systems will further expand and deepen in several asp-ects, including multimodal fusion, high security, high interpretability, low resource consumption, and a comprehensive evaluation system combining large models and automation. Reference | Related Articles | Metrics

Select

Exploration and Outlook on Construction Path of Large Language Models for Trajectory Mining GUO Xuan, YANG Ruiying, CHEN Mingxing, YU Juanxiu, LIU Junnan Computer Engineering and Applications    2026, 62 (18): 1-17.   DOI: 10.3778/j.issn.1002-8331.2512-0420 Abstract （194）      PDF（pc） （2437KB）（297）       Knowledge map    Save Research on large language model (LLM) empowered trajectory mining is conducted to address the limitations of traditional models in cross-domain generalization and multi-task processing. The development status and classification of LLM are reviewed. Architecture optimization, domain fine-tuning, and scene tuning are analyzed to clarify the adaptation mechanisms for improving model adaptability and efficiency. Internal driving forces of LLM-driven trajectory mining are explored from three dimensions：trajectory semantic enhancement, vertical domain reinforcement, and spatio-temporal scenario adaptation. Based on this, a bottom-up architecture is proposed, which includes a general LLM layer, a domain-specific LLM layer, and a scenario-specific LLM layer. Its application value in trajectory data semantic enrichment, domain empowerment, and scenario specialization is elaborated. Finally, challenges such as insufficient multi-source data fusion and lack of domain credibility assurance are summarized, and future trends including cross-modal fusion, visual credibility analysis, and agent collaboration are highlighted. Reference | Related Articles | Metrics

Select

Improved RT-DETR Algorithm for Small Object Detection in UAV Images GUO Jie, HU Jianlong, ZHANG Junchao, WANG Yuxiang, ZHANG Xinyi Computer Engineering and Applications    2026, 62 (8): 93-104.   DOI: 10.3778/j.issn.1002-8331.2507-0309 Abstract （336）      PDF（pc） （2233KB）（294）       Knowledge map    Save To address challenges in UAV aerial imagery, such as small object size, blurred features, and frequent occlusions, an improved real-time detection Transformer (RT-DETR) algorithm is proposed to enhance detection accuracy and robustness in complex scenes. A feature enhancement module, C2f synergistic multi-attention Transformer (C2f_SMT), is designed in the backbone to integrate shallow details and high-level semantics by combining the synergistic multi-attention (SMA) mechanism with Transformer architecture, thereby improving feature representation for small objects. To optimize feature interaction, a dual-attention feature interaction module (DAFI) is introduced, employing a parallel lightweight structure that reduces model complexity while preserving global modeling capacity. To address the issue of small-object information loss during multi-scale fusion, a cross-OmniKernel and small-target preservation feature fusion module (COSPFM) is proposed to enhance multi-scale perception of small targets. Experimental results on the VisDrone2019 dataset show that the improved model achieves a mAP@0.5 of 51.0% and a mAP@0.5:0.95 of 31.5%, outperforming the baseline RT-DETR by 3.1 and 2.2 percentage points, respectively. To evaluate generalization, the model is further tested on the UAVDT dataset, where it demonstrates improved accuracy across various traffic categories (car, truck, bus), achieving a mAP@0.5 of 34.4%, confirming the effectiveness and generalizability of the proposed method in multi-scenario small object detection tasks. Reference | Related Articles | Metrics

Select

PipeSeg-YOLO：Lightweight Instance Segmentation Algorithm for Urban Drainage Pipeline Defects SU Wenxiang, XU Junsong, ZHANG Hui, LUO Min Computer Engineering and Applications    2026, 62 (9): 307-324.   DOI: 10.3778/j.issn.1002-8331.2509-0311 Abstract （259）      PDF（pc） （159509KB）（280）       Knowledge map    Save To address the challenges of harsh imaging environments, diverse defect morphologies within urban drainage pipelines, and the urgent demand for lightweight, real-time instance segmentation algorithms for inspection robots, this paper proposes an efficient and accurate instance segmentation algorithm named PipeSeg-YOLO. Built upon the YOLOv8n-seg baseline, the algorithm achieves systematic lightweight optimization through the synergistic design of three core modules：(1) An efficient and adaptive backbone network (C2f-DHC) based on dynamic harmonic convolution, which enhances the feature representation capability for diverse defect morphologies with extremely high parameter efficiency by re-parameterizing dynamic convolution in the frequency domain; (2) A GD2-PAN neck built upon the advanced “gather-and-distribute”topology, which integrates efficient large-kernel convolution, a gated dynamic fusion mechanism, and the Slim-Neck hardware-aware paradigm to systematically enhance the efficiency and fidelity of multi-scale feature fusion; (3) A novel segmentation head (HLR-Head) based on category-adaptive low-rank reconstruction, which effectively resolves the computational bottleneck of mask generation. Experimental results demonstrate that, compared to the YOLOv8n-seg baseline, PipeSeg-YOLO reduces parameters and GFLOPs by 44.1% and 46.3%, respectively, while simultaneously improving segmentation accuracy ([mAPmask]@0.5) by 3.7 percentage points. Its inference speed reaches 208 FPS. Deployment verification on the NVIDIA Jetson Orin Nano edge device achieves an inference speed of 50 FPS, demonstrating the model’s feasibility for efficient deployment in real-world, resource-constrained scenarios and providing an effective solution for edge computing applications in pipeline defect detection and segmentation. Reference | Related Articles | Metrics

Select

Survey of Vision-Language-Action Models： Towards General-Purpose Robots CHEN Wenqi, CHEN Jiafeng, ZHI Pengxiang, SHI Lulu, WEN Luhong Computer Engineering and Applications    2026, 62 (11): 17-40.   DOI: 10.3778/j.issn.1002-8331.2511-0018 Abstract （370）      PDF（pc） （1516KB）（267）       Knowledge map    Save The vision-language-action (VLA) model, as an important research direction in embodied intelligence, aims to overcome the semantic fragmentation and generalization bottleneck of traditional robot systems in open environments by deeply integrating perception, understanding, and action. Based on the relevant progress in the VLA field, this paper first reviews the development history and current research status of VLA models. Then, from the perspective of system architecture, existing models are classified into three types： monolithic, cascaded, and hierarchical, and the design ideas and representative progress of each type of architecture are analyzed in depth. At the same time, the relevant datasets and evaluation benchmarks that support VLA research are systematically summarized. Finally, based on the representative achievements in recent years, the core challenges faced by current VLA models are analyzed, and the future development directions are prospected. Reference | Related Articles | Metrics

Select

Survey on Large Model-Assisted Reward Design Methods for Reinforcement Learning CAO Yuqing, CHEN Xiliang, DONG Haoyang, ZHOU Xin, SUN Mingwei Computer Engineering and Applications    2026, 62 (9): 83-107.   DOI: 10.3778/j.issn.1002-8331.2509-0180 Abstract （258）      PDF（pc） （1374KB）（261）       Knowledge map    Save Reward design stands as a core challenge in reinforcement learning. Traditional methods, which rely on experts’ manual design, are plagued by several limitations：strong subjectivity, difficulty in expressing complex objectives, high susceptibility to reward hacking, and the sparse rewards problem. Leveraging their robust capabilities in semantic understanding, task decomposition, and code generation, large models have enabled end-to-end automated mapping from “human intentions” to “executable reward functions”. This significantly lowers the barrier to reward design and overcomes the bottlenecks of traditional methods, which rely heavily on expert knowledge and suffer from ambiguities in goal expression. This paper systematically reviews the research progress of LLM-assisted reward design in reinforcement learning. It analyzes the limitations of traditional reward design methods. It proposes a classification framework based on the roles of LLMs in reward design, namely reward function initializer, “human-in-the-loop” iterator, and “human-out-of-the-loop” iterator; it also elaborates on the representative works, underlying mechanisms, advantages, and disadvantages of each category of methods. It expounds on the fundamental framework of reinforcement learning and the challenges of reward design, analyzing the intrinsic mechanisms and value of large model assistance. It reviews the limitations of traditional reward design methods, constructs a “Reward Initializer-Reward Iterator” classification framework, and details the representative work, mechanisms, and applicable scenarios of various methods. Furthermore, it summarizes the key challenges faced by current research at the “Model-Methodology-System” level in terms of model reliability, interaction efficiency, and engineering safety, and proposes corresponding technical pathways. Reference | Related Articles | Metrics

Select

PCSED-YOLO： Study on Cross-Scale Multi-Object Wearable Detection Algorithm in Complex Environments XUE Guanghui, YAN Zhaoyang, WU Mian Computer Engineering and Applications    2026, 62 (5): 88-105.   DOI: 10.3778/j.issn.1002-8331.2505-0210 Abstract （235）      PDF（pc） （8315KB）（257）       Knowledge map    Save Non-standard safety wear by workshop workers during operations can affect physical health and lead to safety accidents, including injuries or fatalities. Computer vision-based detection of personal protective equipment has become a research focus. Workshops have complex environments with numerous devices and harsh conditions, dust generated during production blurs or obscures targets, presenting challenges for multi-target detection across scales. Existing algorithms struggle with low accuracy and high false negative rates for small targets like masks. To address this, an improved YOLO-based detection algorithm, PCSED-YOLO, is proposed. The PPA module enhances feature extraction for small targets and multi-scale detection. The CAFM module in C2PSA achieves local and global feature complementarity, enhancing recognition in dusty environments. SPD-Conv replaces original convolutional layers for lossless down-sampling, improving detection of small, low-resolution targets. The SEv2 improvement of the spatial pyramid pooling layer enhances global context control and feature extraction for multi-class, cross-scale targets. Dynamic-Conv in the detection head adjusts convolution kernel size and shape for precise cross-scale detection. A higher resolution P2 detection layer improves small target accuracy. A dataset of workers’ protective wear is compiled for ablation and comparative experiments. Results show that PCSED-YOLO excels in handling targets of various sizes, achieving a mAP@0.5 of 0.946, a 0.077 improvement over the baseline. AP@0.5mask reaches 0.887, a 0.236 improvement, AP@0.5no-helmet is increased by 0.037 to 0.958, and AP@0.5vest is improved by 0.006 to 0.991. F1-Score and P-R curve metrics show significant enhancement. Compared to advanced models, PCSED-YOLO achieves the best accuracy on the dataset, demonstrating strong capabilities in complex environments and providing a new algorithm for multi-target detection of protective wear. Reference | Related Articles | Metrics

Select

Review on Machine Learning Modeling and Optimization of Battery Thermal Management Systems for Electric Vehicles GUO Hongfei, CUI Yu, ZHANG Rui Computer Engineering and Applications    2026, 62 (16): 1-20.   DOI: 10.3778/j.issn.1002-8331.2509-0387 Abstract （310）      PDF（pc） （1225KB）（254）       Knowledge map    Save The battery thermal management system (BTMS) is a key subsystem that ensures the safe operation, energy efficiency, and longevity of traction batteries in electric vehicles. Traditional physics-based methods rely on fine-grained modeling and parameter tuning, and under complex operating conditions they often suffer from model mismatch, limited real-time performance, and high computational cost. With advances in sensing and computing, machine learning has been widely applied to thermal-state modeling, thermal anomaly detection, and cooling-control optimization, and has also driven the adoption of lightweight techniques, such as model compression, quantization, and distillation on in-vehicle embedded platforms. This paper systematically reviews related progress of machine learning in thermal state modeling and prediction：Supervised learning shows clear advantages in temperature prediction, unsupervised and semi-supervised learning perform well in thermal anomaly detection, and reinforcement learning exhibits adaptive strengths in optimizing cooling-control strategies. By comparing trade-offs among accuracy, real-time capability, and computational demand, this paper emphasizes that physics-data integration and lightweight implementation are pivotal for engineering deployability. Overall, machine learning complements traditional approaches, improves prediction and control performance, and provides guidance for the intelligent development and engineering application of BTMS. Reference | Related Articles | Metrics

Select

MCT-YOLO: Small Object Detection Algorithm for Aerial Imagery ZHOU Zhongkun, ZHANG Yang, ZHANG Yu, MENG Ke, YUAN Zhaoyang Computer Engineering and Applications    2026, 62 (8): 130-141.   DOI: 10.3778/j.issn.1002-8331.2509-0032 Abstract （317）      PDF（pc） （2760KB）（251）       Knowledge map    Save To address the issue of low detection accuracy caused by small scale, low resolution, and high-density distribution of image targets in unmanned aerial vehicle-assisted intelligent transportation scenarios, an improved YOLOv11n-based algorithm for small object detection from an aerial perspective, named MCT-YOLO, is proposed. The detection layer structure is optimized by specifically adding a very small object detection head and removing the large object detection head to capture more local information and reduce feature interference, thereby adapting to small object detection tasks. A multi-frequency interactive downsampling (MFID) module is proposed to minimize feature loss during downsampling and retain more image information. The MobiVari structure is introduced to build the C3k2_MV module for deep feature extraction, enhancing detailed feature representation. Additionally, a multi-scale information fusion method called triple complementary fusion (TCF) is designed to integrate contextual information of different semantics, achieving complementary semantic and detailed information to improve small object detection capability. The proposed algorithm achieves mAP50 and mAP50-95 scores of 40.9% and 25.0%, respectively, on the VisDrone2019-DET dataset, representing improvements of 7.7 and 5.8 percentage points over the baseline network YOLOv11n. Meanwhile, the algorithm??s parameter count is reduced by 7.6%, making it suitable for small object detection scenarios in drone aerial imagery. Reference | Related Articles | Metrics

Select

Research Progress and Key Technologies of Bionic Behavior Control for UAV Swarms XIA Hengyu, HE Ming, HAN Wei, WU Jingjing, ZHAO Xiaojun Computer Engineering and Applications    2026, 62 (17): 1-17.   DOI: 10.3778/j.issn.1002-8331.2512-0142 Abstract （251）      PDF（pc） （2058KB）（249）       Knowledge map    Save As the tasks and environments faced by unmanned aerial vehicle(UAV) swarms become increasingly complex, the shortcomings of traditional control methods in terms of adaptability and robustness have become more prominent. Mapping the behavioral mechanisms of biological groups to unmanned aerial vehicle(UAV) swarm systems has become an important research direction. However, existing research lacks the refinement of the behavioral mechanisms of different groups and the systematic review of related technologies. For this reason, this paper systematically reviews the current research status of bionic behavior control for UAVs, with a focus on three typical objects：bee colonies, wolf packs, and bird flocks. Firstly, the paper expounds the internal structure, interaction methods and behavioral mechanisms of three typical bionic groups. Secondly, it analyzes the key technologies of bionic behavior control for UAVs and summarizes the advantages and disadvantages of each method as well as their applicable scenarios. To address the challenges of adaptability and flexibility in bionic behavior control for unmanned aerial vehicle swarms, a bionic behavior control strategy based on behavior cloning is proposed. Finally, the paper summarizes and looks forward to the future development trends of the research on bionic behavior control methods for unmanned aerial vehicle swarms. Reference | Related Articles | Metrics

Select

Survey of Privacy-Preserving Techniques in Federated Generative Adversarial Networks HUO Zheng, WANG Suzhen, ZHANG Tengfei Computer Engineering and Applications    2026, 62 (12): 1-18.   DOI: 10.3778/j.issn.1002-8331.2508-0265 Abstract （280）      PDF（pc） （1102KB）（249）       Knowledge map    Save Federated generative adversarial networks (FedGAN) have emerged as a promising privacy-preserving framework by integrating generative modeling with federated learning. Unlike traditional centralized GAN, FedGAN enables joint model training without exposing raw data, thereby reducing the risk of privacy leakage. However, due to the intrinsic capability of generative models to reconstruct data distributions, FedGAN remains vulnerable to various privacy attacks such as membership inference and model extraction. This paper provides a comprehensive overview of privacy attack models in FedGAN and surveys the research progress of differential privacy mechanisms within this framework. Specifically, it analyzes the architectural characteristics of FedGAN and the implementation of mainstream privacy-preserving techniques, including differentially private stochastic gradient descent (DPSGD) and Rényi differential privacy (RDP). The paper also briefly introduces the recent applications of knowledge distillation, homomorphic encryption, and secures multi-party computation in the FedGAN context. Finally, this paper summarizes the current challenges and discusses potential future research directions. Reference | Related Articles | Metrics

Select

Review of Research on Underwater Environment and Underwater Robot Modeling Based on Virtual Simulation Platforms ZHOU Huanyin, LI Wentao, LIU Yijia, HUANG Jiahao Computer Engineering and Applications    2026, 62 (14): 1-16.   DOI: 10.3778/j.issn.1002-8331.2507-0205 Abstract （278）      PDF（pc） （1076KB）（248）       Knowledge map    Save Underwater robot research and validation face significant challenges due to the complexity of the environment, the high cost of physical testing, and potential safety risks. Virtual simulation platforms address these issues by modeling both underwater environments and robotic systems, enabling efficient performance verification and optimization under near-realistic conditions. This approach greatly reduces development costs and risks. This paper reviews the functional features and integration mechanisms of mainstream underwater robot simulation platforms. It analyzes key environment modeling techniques, including hydrodynamics, underwater light propagation, and three-dimensional terrain and obstacle representation. Advances in robot modeling, motion control algorithms, path planning, and obstacle avoidance are also examined. Moreover, the study highlights cooperative simulation mechanisms that integrate environment and robot modeling. These mechanisms, supported by closed-loop architectures and real-time bidirectional data exchange, allow for high-fidelity physical interactions between robots and their surroundings. The limitations of current platforms are summarized. The review suggests that future developments will move toward higher fidelity, broader generalization, and deeper integration, providing a stronger foundation for underwater robot research and applications. Reference | Related Articles | Metrics

Select

Improved YOLOv11 Algorithm for Aerial Small Target Detection in UAVs JING Tingting, CAO Yudong, CHEN Xin, WANG Dongxia Computer Engineering and Applications    2026, 62 (2): 138-148.   DOI: 10.3778/j.issn.1002-8331.2506-0186 Abstract （305）      PDF（pc） （2738KB）（243）       Knowledge map    Save In the field of unmanned aerial vehicle aerial photography, the proportion of small targets is high, and the resources on board are limited, which results in low precision of existing detection models and difficulty in achieving a good balance between detection performance and resource consumption. To alleviate these issues, an improved YOLOv11 model is proposed. Firstly, the MDPE module is constructed as the attention module of the backbone network. By introducing a dynamic channel splitting strategy and a lightweight smoothing network, it achieves adaptive allocation of feature resources and enhancement of local details. Secondly, the DAFI module is designed to replace the feature fusion module in the neck network. By combining the feature enhancement module with a dual dynamic weighting mechanism, it enhances the algorithm’s perception of the spatial distribution of targets. Thirdly, block-based dynamic alignment convolution is embedded in the regression branch of the detection head to achieve adaptive alignment and fusion of feature maps, thereby improving the model’s detection capability for targets of different scales. Finally, by introducing the aspect ratio difference term and combining it with the small target weight adjustment mechanism, the ARSIoU loss function is designed to improve the model’s localization accuracy and bounding box regression efficiency for small targets. Experiments on the VisDrone2019 dataset demonstrate that the improved model outperforms the original YOLOv11 model in terms of precision, recall, mAP@50 by 3.5, 1.9, 3.2 percentage points respectively, indicating superior comprehensive performance compared to other small target detection algorithms tested, making it suitable for unmanned aerial vehicle aerial small target detection applications in complex scenarios. Reference | Related Articles | Metrics

Select

Research on Low-light Object Detection Method of Improved YOLOv11 LI Junlin, ZHANG Xuesong, SONG Cunli, LI Guangyu Computer Engineering and Applications    2026, 62 (5): 106-119.   DOI: 10.3778/j.issn.1002-8331.2506-0263 Abstract （228）      PDF（pc） （3198KB）（236）       Knowledge map    Save For the performance bottleneck in detecting low-light complex scenarios, this paper proposes a low light target detection algorithm ELS-YOLO (enhanced low-light scene-YOLO) based on improved YOLOv11. Firstly, EfficientNetV2 is introduced to replace the backbone network, simplifying the model structure while enhancing the complex interaction between local details and global background. Secondly, a frequency domain perception module is proposed to further improve the network’s perception of image details, helping the model better identify the contours and dark details of objects under low-light conditions. Finally, a loss function based on the wise stepwise gain allocation strategy is designed, which continuously improves detection accuracy while enhancing the adaptability to light and noise changes. The ablation and comparative experiments are carried out on the ExDark, NOD (night object detection) and the VOC2012 datasets. The results demonstrate that on the three datasets, the proposed algorithm outperforms the baseline model by 4.2 and 2.6 percentage points, 2.4 and 0.7 percentage points, and 0.7 and 0.3 percentage points in terms of mAP@0.5 and mAP@0.5：0.95, respectively. These findings validate the effectiveness of the algorithm in low-light object detection scenarios. Reference | Related Articles | Metrics

Select

Insulator Defect Detection Based on Multi-Scale Features and Channel Perception JI Kai, ZHANG Wenbin, CHEN Xiaotao Computer Engineering and Applications    2025, 61 (24): 240-250.   DOI: 10.3778/j.issn.1002-8331.2409-0265 Abstract （140）      PDF（pc） （70002KB）（233）       Knowledge map    Save In order to solve the problems of complex background, diverse defect types and large-scale differences in insulator images in UAV aerial photography, an insulator defect detection algorithm based on multi-scale features and channel perception is proposed. The feature maps at different levels are introduced into the multi-scale feature fusion module to combine the multi-scale feature information, and a series of parallel deep convolutions are used to fully extract the context information. The adaptive channel-aware down sampling module is embedded in the backbone network, the spatial dimension of the input features is converted into the channel dimension to realize down sampling, so as to ensure that the feature map avoids the loss of small-scale feature information in the down sampling process, and the efficient channel attention mechanism is used to improve the model’s perception ability of important channel information. On the basis of adding a small-scale object detection layer, a lightweight shared ghost convolutional detection head is designed to improve the model’s detection ability of multi-scale targets under the condition of reducing the number of model parameters. Experimental results show that the proposed model has achieved good results for different types and multi-scale insulator defects, and its average detection accuracy is 91.4%, which is 5.2 percentage points higher than that of the benchmark model, and the overall performance is better than that of other insulator defect detection algorithms. Reference | Related Articles | Metrics

Select

DRGF-YOLO：Multi-Scale Feature Fusion for Safety Detection in Chemical Plants CHANG Ligang, XU Binhan, SHEN Liangchang, DU Yanhui Computer Engineering and Applications    2026, 62 (12): 93-105.   DOI: 10.3778/j.issn.1002-8331.2509-0233 Abstract （246）      PDF（pc） （4390KB）（223）       Knowledge map    Save Chemical plants involve complex environments with multiple unsafe behaviors and potential hazards that, if not detected in time, may lead to severe accidents such as fires and explosions. To address the challenges of small objects, distant targets, and complex backgrounds, this paper proposes DRGF-YOLO, an improved YOLOv13 model. The backbone incorporates an MGS-Block to enhance multi-scale feature extraction, while an MGCA module strengthens channel modeling and robustness. In the neck, a DR-GFPN structure achieves dynamic cross-layer and cross-scale fusion, effectively improving small-object representation. To alleviate data scarcity, a CHEM-SAFE-8 dataset is constructed, combining Internet images and real surveillance data from chemical plants. Experimental results demonstrate that DRGF-YOLO achieves higher mAP, Precision, and Recall than YOLOv13-L and other baselines on the CHEM-SAFE-8-Plant dataset, especially for small objects and complex scenarios, while maintaining real-time efficiency. Additional tests on public datasets such as D-Fire and VisDrone2019 further validate the superior performance and generalization ability of the proposed model compared with recent improved models. Reference | Related Articles | Metrics

Select

HM-YOLO: Lightweight Aerial Image Detection Algorithm Fusing Multi-Scale Features LI Jun, DING Binbin, SHI Weijuan, YANG Lin Computer Engineering and Applications    2026, 62 (1): 87-100.   DOI: 10.3778/j.issn.1002-8331.2504-0309 Abstract （239）      PDF（pc） （5025KB）（222）       Knowledge map    Save Aiming at the problems of tiny target size, complex background environment and difficult feature extraction in the UAV aerial image detection task, a lightweight aerial image detection algorithm HM-YOLO is proposed on the basis of YOLOv11. First, downsampling and upsampling of features of different scales in the backbone network extends the information interaction between different feature channels, and at the same time optimizes the shallow feature graph size to adapt to the tiny targets in aerial images; second, an efficient feature extraction module C3k2_MSEE is designed, which first divides the feature graph using adaptive mean pooling, and then highlights the edge information through the edge enhancement module to avoid the loss of feature information of the small targets in the deep network; then, a hierarchical attention fusion module HAFB is proposed, which strengthens the model’s understanding of the contextual contextualization of the features in the backbone network by constructing a local and global dual-channel attention network, which strengthens the model’s ability to integrate contextual information; finally, DyHead, a dynamic detection head with multiple attention mechanisms, is introduced to further optimize the ability to perceive small target feature information. And HM-YOLO is lightweighted using the LAMP pruning method and BCKD knowledge distillation strategy, which significantly compressees the volume of the model. Experimental results on the Visdrone2019 dataset demonstrate that the improved algorithm achieves 8.4, 5.7, and 8.4 percentage points increases in accuracy, recall, and mAP@50, respectively, are able to effectively cope with the challenges in the task of target detection in UAV aerial images. Reference | Related Articles | Metrics

Select

Survey of Vision Transformers for Fine-Grained Image Classification WEN Shixiong, ZHI Min Computer Engineering and Applications    2025, 61 (23): 24-37.   DOI: 10.3778/j.issn.1002-8331.2503-0014 Abstract （321）      PDF（pc） （2320KB）（221）       Knowledge map    Save Fine-grained image classification (FGIC) aims to identify subcategories that are visually highly similar yet exhibit subtle differences. With the rapid advancement of deep learning, FGIC algorithms have gradually evolved from traditional fully supervised learning to weakly supervised approaches. Vision Transformers (ViTs), leveraging multi-head self-attention mechanisms, eliminate the reliance on manual annotations and overcome the limitations of convolutional neural networks (CNNs) in terms of receptive field size and global modeling capacity, becoming one of the mainstream methods for this task. This paper first outlines the key characteristics and challenges of FGIC, and briefly introduces the architecture and advantages of ViT. Based on different feature fusion strategies, existing ViT-based improvements are categorized into hierarchical fusion, multi-local fusion, and multi-granularity fusion. The modifications of each category are illustrated in detail, and their underlying mechanisms are systematically analyzed and summarized. In addition, commonly used public datasets are reviewed, and future research directions are proposed based on current limitations, aiming to further explore the potential of ViT in FGIC tasks. Reference | Related Articles | Metrics

Select

Research on Framework of Global Airshow Knowledge Graph Construction and Question-Answering System Based on Large Language Models CAI Jintian, HAO Yaohui, CHEN Xin, QIN Yudong Computer Engineering and Applications    2026, 62 (9): 108-121.   DOI: 10.3778/j.issn.1002-8331.2507-0068 Abstract （262）      PDF（pc） （3551KB）（220）       Knowledge map    Save Against the backdrop of rapid advancement of global aerospace technology, airshows serve as a pivotal platform for international industry exchanges and technological demonstrations, accumulating massive heterogeneous information including airshow history, exhibit technical parameters, and social media public opinion on airshow topics. Efficient information acquisition and utilization through existing technologies hold significant practical value. A global airshow domain knowledge graph construction method based on large-model research is proposed. By integrating retrieval augmentation generation (RAG) technology, an airshow domain Q&A system framework is constructed on this graph through fine-tuning low-rank adaptation (LoRA) and implementing chain-of-thought design for pre-trained models, constructing the domain knowledge graph, and introducing document parsing, semantic chunking, hybrid retrieval, and cue enhancement techniques. Dataset-based tests demonstrate that the proposed method improves the benchmark model in accuracy, precision, and F1-score metrics, and subjective and objective evaluations of the Q&A system show that the RAG framework effectively reduces model “hallucination” frequency and enhances professional Q&A capabilities in specialized domains. Reference | Related Articles | Metrics

Select

Survey of Large Language Model Driven Intelligent Processing of Tabular Data WANG Xingkai, XI Xuefeng, CUI Zhiming, WANG Fei, ZHENG Qian Computer Engineering and Applications    2026, 62 (9): 20-45.   DOI: 10.3778/j.issn.1002-8331.2506-0198 Abstract （294）      PDF（pc） （2442KB）（220）       Knowledge map    Save Due to the structured and semi-structured characteristics of tabular data, it plays an irreplaceable pivotal role in tasks such as data warehouse construction, knowledge graph alignment and feature engineering. In recent years, large language models provide new technical possibilities for intelligent processing of “unified modeling” tables by means of extremely long context windows, zero/few-shot transfer capabilities, and pluggable tool interfaces. Firstly, this paper systematically combs the mainstream representation forms of tabular data and their influence on the complexity of subsequent tasks, and establishes the mapping between task pedigrees and typical datasets covering four categories of tasks：structure understanding, data understanding, reasoning, and data augmentation. From four dimensions of pre-training strategy, text serialization and embedding, prompt engineering and inference chain design, and retrieval enhancement generation and dynamic feedback mechanism, the key technologies of large language models enabling intelligent processing of forms are summarized. Secondly, according to the four modules of structure parsing, table understanding, reasoning generation and data governance, the latest applications of large language models in table detection and positioning, entity alignment, Text-to-SQL and other tasks are analyzed. Finally, the core bottlenecks faced by large language models in tabular scenarios are pointed out：insufficient structural understanding, cross-modal alignment errors, scarcity of labeled data, and poor execution reliability. Looking forward to the future, with the continuous evolution of the multi-agent collaborative decision-making framework, the structure-symbol hybrid reasoning paradigm, and the industry-specific instruction fine-tuning technology, the large language model is expected to complete the intelligent processing tasks of forms robustly, interpretable, and efficient in real business scenarios. Reference | Related Articles | Metrics

Select

Improving Water Debris Detection Method of YOLOv12 ZHU Yufan, YANG Jikai, LI Wei Computer Engineering and Applications    2026, 62 (1): 101-111.   DOI: 10.3778/j.issn.1002-8331.2505-0249 Abstract （213）      PDF（pc） （28047KB）（218）       Knowledge map    Save To address the issue of outdated models, high rates of missed and false detections, and the inability of existing water surface debris detection methods to meet the requirements of efficient water surface environmental management, this paper proposes an improved water debris detection method, WDW-YOLO, based on YOLOv12. This method introduces WTConv based on wavelet transform to improve the C3k2 module in the neck network, designing the C3k2_WTConv module. This design effectively expands the model’s receptive field while maintaining controllable model parameters. For the C3k2 module in the backbone network, it introduces DynamicConv to innovate and obtain the C3k2_DynamicConv module. This approach increases the number of learnable parameters while limiting the growth of computational load, thereby enhancing the model’s ability to capture feature information and improving detection accuracy. Additionally, it replaces CIoU with WiseIoUv3 and explore the optimal combination of hyperparameters to balance samples with different detection difficulties and enhance the model’s ability to detect small objects. Experiments are conducted on the FloW_IMG dataset, including comparative and ablation studies. The results show that the proposed WDW-YOLO model achieves a 2.7 and 1.8 percentage points increase in mAP50 and mAP50-95, respectively, compared to the baseline YOLOv12n model, reaching 90.7% and 50.6%. This improvement is achieved while reducing model complexity and enhancing detection accuracy. The WDW-YOLO model demonstrates high reliability and practicality for automated water surface cleaning and water body protection. Reference | Related Articles | Metrics

Select

Research Progress and Prospects of Few-Shot Object Detection in Remote Sensing Images GAO Guangshuai, ZHANG Zhilin, DONG Yan Computer Engineering and Applications    2026, 62 (16): 21-41.   DOI: 10.3778/j.issn.1002-8331.2510-0057 Abstract （231）      PDF（pc） （1940KB）（217）       Knowledge map    Save This paper systematically summarizes recent advances of few-shot object detection in optical remote sensing images, and analyzes key challenges and corresponding solutions in the field. Few-shot object detection in remote sensing images aims to accurately detect novel object categories using only a few annotated samples, making it suitable for data-scarce scenarios in remote sensing images. The paper outlines the fundamental definition and framework of few-shot object detection in remote sensing images, and addresses six major challenges：scale variation, shape diversity, background interference, class imbalance, incomplete annotation, and orientation uncertainty. Various deep learning-based methods is reviewed, including multi-scale feature fusion, intra/inter-class variation handling, background suppression, and Pseudo-label generation strategy. Commonly used datasets and evaluation metrics are introduced, and representative algorithms are compared to identify core techniques affecting performance. Finally, future research directions, such as self-supervised learning, multimodal fusion, and incremental learning, are discussed, providing valuable references for research on few-shot object detection in remote sensing images. Reference | Related Articles | Metrics

Select

Review of Path Planning Algorithms for Mobile Robots ZHANG Yonghong, GUO Zijian, LU Zhuheng, JIANG Liang, CAO Haixiao Computer Engineering and Applications    2026, 62 (2): 26-39.   DOI: 10.3778/j.issn.1002-8331.2507-0291 Abstract （339）      PDF（pc） （1715KB）（212）       Knowledge map    Save Path planning algorithms are one of the key technologies for enabling autonomous navigation of mobile robots, and their performance determines the quality of path planning. To comprehensively understand the research status and development of mobile robot path planning algorithms, a systematic review of commonly used algorithms is conducted. Firstly, based on the characteristics of path planning algorithms, they are classified into traditional algorithms, sampling-based algorithms, artificial intelligence-based algorithms, and intelligent-biologically-inspired algorithms. Secondly, based on this classification, the algorithm principles and practical application scenarios are briefly introduced, with a focus on the research achievements of various algorithms in recent years, and the advantages and disadvantages of each algorithm are summarized and compared. Then, four algorithms are selected to verify the effectiveness of the algorithms in the same simulation environment. Finally, the future development trends of mobile robots are prospected to provide a reference for the research on mobile robot path planning. Reference | Related Articles | Metrics

Select

Survey of Neural Network Filter Pruning Technology WANG Lin, SONG Quanrun, GENG Shichao, LUAN Zhongzhi Computer Engineering and Applications    2026, 62 (2): 1-25.   DOI: 10.3778/j.issn.1002-8331.2503-0280 Abstract （263）      PDF（pc） （1413KB）（211）       Knowledge map    Save With the improvement of software and hardware resources and computing capabilities, deep neural networks have rapidly developed in various fields such as computer vision, natural language processing, and image generation, leading to breakthroughs in areas such as autonomous driving and medical diagnosis through deep learning. However, as the depth of the model increases, the large number of parameters and the consumption of computing resources cause the model to become too complex to train and deploy in resource-constrained environments. To reduce the complexity of network models and improve their efficiency, researchers have proposed pruning methods to compress and accelerate models by reducing redundant parameters and connections. Filter pruning is one of the most important methods to optimize convolutional neural networks. It mainly changes the number of filter groups and feature channels to accelerate the network and does not depend on specific algorithm or hardware platform. So the paper reviews the research progress of filter pruning technology at home and abroad in recent years, classifies and summarizes the filter importance evaluation, pruning and fine-tuning mode design, and summarizes the experiments of the mainstream filter pruning methods to analyze the influence of filter pruning on model accuracy and parameter number, and also discusses the future research direction. Reference | Related Articles | Metrics