HUANG Wenlong , ZHAO Haohao , KANG Jian , ZHI Xiaodong , WANG Dongchuan , ZHOU Weixun , NI Huan , GUAN Haiyan
2025, 17(2):151-164. DOI: 10.13878/j.cnki.jnuist.20241009001
Abstract:To accurately detect road damages with large size differences and small scales in vehicle-mounted images,this paper presents a real-time road damage detection model based on improved YOLOv5s,termed as VRD-YOLO (Vehicle-mounted image Road Damage Detection-YOLO).Firstly,a Channel Mix Slide Transformer (CMST) module is proposed to enhance the model's global context modeling capability and strengthen the extraction of fine-grained road damage semantic feature information.Secondly,a generalized feature pyramid with cross-layer fusion and cross-scale fusion is introduced to expand the network receptive field and strengthen the fusion of multi-scale damage features.Thirdly,to optimize the model's feature response and further improve detection performance,a dynamic detection head is designed to achieve scale perception,spatial perception,and task perception.Finally,a Vehicle-mounted Image Road Damage Dataset (VIRDD) is constructed to expand the quantity and types of existing road damage datasets,and ablation and comparative experiments are conducted based on this dataset.Experimental results show that the VRD-YOLO achieves a mean Average Precision (mAP@0.5) of 74.45% on the VIRDD dataset,with a detection speed reaching 28.56 frames per second.Compared to the YOLOv5s model,VRD-YOLO improves the precision,recall,F1 score,and mAP by 2.79,2.32,2.54,and 3.19 percentage points,respectively.Additionally,compared with six other classical and mainstream object detection models,the proposed VRD-YOLO attains the highest detection accuracy with the smallest model parameter count of 9.68 million,verifying its effectiveness and superiority.
LI Guoli , CHEN Yanming , XIA Jiakang , ZOU Xincan
2025, 17(2):165-171. DOI: 10.13878/j.cnki.jnuist.20241024001
Abstract:To address the deficiencies of traditional graph neural network methods in point cloud semantic segmentation,such as high requirements for supervision accuracy,one-way only node label propagation,and neglect of global information,this paper proposes a point cloud semantic segmentation method based on bidirectional attention mechanism.Firstly,the point cloud is over-segmented into superpoints and a superpoint graph is constructed,thus introducing the point cloud classification problem into the superpoint graph network framework.Subsequently,the two-way attention module is utilized to alternately focus on superpoints and update their features according to the weights of neighboring superpoints,enabling the two-way information propagation.Unlike previous graph pooling methods,this study applies both maximum pooling and average pooling,and combines their pooled features.Finally,the public dataset Semantic 3D is used for training and experiments.The results show that the proposed method can effectively correct labelling errors while coupling local features with long-range information,and the mean Intersection over Union (mIoU) and overall Accuracy (oAcc) of the dataset are 75.4% and 95.1%,respectively,exhibiting a better label delivery mechanism and higher classification accuracy compared with existing methods.
TANG Feifei , YANG Hao , LIU Na , JIANG Min , PANG Rong , ZHANG Peng , ZHOU Zelin
2025, 17(2):172-180. DOI: 10.13878/j.cnki.jnuist.20240927001
Abstract:To tackle the current challenges of low efficiency,poor performance,and inadequate real-time capabilities in bridge crack detection,this paper introduces a drone-based image detection method for bridge cracks using an improved YOLOv8 model.Firstly,the dynamic snake convolution kernel is integrated into the C2f module in the backbone of YOLOv8 to enhance the crack feature extraction.Then,the Context Augmentation Module (CAM) is introduced to improve the detection capability for small targets.Finally,the influence of low-quality datasets on detection results is reduced via optimizing the prediction box loss function.Experimental results show that the improved model achieves a GFLOPs of 14.4 and a mean Average Precision (mAP@50) of 94%,exhibiting a significant accuracy improvement compared to the baseline models.The detection speed reaches 147 frames per second,satisfying the requirements for real-time crack detection by UAVs.
ZHANG Yuting , ZHENG Dehua , LI Siyuan
2025, 17(2):181-190. DOI: 10.13878/j.cnki.jnuist.20240513002
Abstract:Aiming at the extraction of cavern surface deformation from three-dimensional laser scanning dense point clouds,we propose a method integrating the Multiscale Model-to-Model Cloud Comparison (M3C2) with an improved Alpha Shapes algorithm.First,the two-phase surface point cloud data are registered,and the improved Alpha Shapes algorithm is used to identify the outer contour point clouds.After the fine registration of these two-phase outer contour point clouds,the M3C2 algorithm calculates the deformation value of each point,and finally the continuous deformation regions are extracted through distance clustering.Experimental results show that the proposed method effectively eliminates the points at small furrows as well as those affected by mixed pixels.Specifically,the removal rates of point clouds in the two phases within 10 m from the scanner to the cavern section are 14.17% and 13.52%,respectively,which are 6.25% and 6.42% within 70 m.This method accurately and efficiently extracts the cavern surface deformation regions with more than twice the registration error.
2025, 17(2):191-202. DOI: 10.13878/j.cnki.jnuist.20240229001
Abstract:To address the low visual quality of stego-image in existing color image Reversible Data Hiding (RDH) algorithms,a novel RDH scheme utilizing multi-level interpolation prediction and global sorting is proposed.Firstly,to fully exploit the features of different texture regions in the image,a multi-level interpolation prediction method is designed to significantly improve the prediction accuracy of pixels.Then,a complexity-based global sorting strategy is designed to sort the prediction errors in the three channels of color images respectively,thereby fully utilizing the global characteristics of the prediction errors in each channel to generate a more concentrated Three-Dimensional Prediction Error Histogram (3D PEH).Finally,an adaptive 3D mapping strategy is used to modify the error histogram and embed secret data.Experimental results show that the proposed approach outperforms some of the latest schemes in embedding performance.
CUI Shaojun , JI Fanfan , WANG Ting , YUAN Xiaotong
2025, 17(2):203-214. DOI: 10.13878/j.cnki.jnuist.20230927001
Abstract:Here,we propose a pruning and optimization approach based on Gradient Weight Pursuit (GWP) to address the overfitting in unsupervised domain,which manifests as significantly lower accuracy on downstream tasks compared to that on training sets.To tackle the overfitting challenge in unsupervised domain,we employ the dense-sparse-dense strategy,focusing on both difference-based and adversarial adaptive methods.First,the network is pretrained intensively to identify crucial connections.Second,during the pruning stage,the optimization algorithm in this paper distinguishes itself from original dense-sparse-dense strategy by jointly considering both weight and gradient information.Specifically,it leverages both weight (i.e.zero-order information) and gradient (i.e.first-order information) to influence pruning process.In the final dense phase,the pruned connections are restored and the dense network is retrained with a reduced learning rate.Finally,the obtained network achieves desirable outcomes in downstream tasks.The experimental results show that the proposed GWP approach can effectively improve the accuracy of downstream tasks,offering a plug-and-play capability compared with original difference-based and adversarial domain adaptation methods.
2025, 17(2):215-226. DOI: 10.13878/j.cnki.jnuist.20230810003
Abstract:As a crucial component of weather systems,3D cloud simulation plays a significant role in various fields such as military and aviation.However,the current mainstream Bounding Volume Hierarchy (BVH) algorithm exhibits inefficient rendering performance when dealing with non-uniform and large-volume clouds.Here,a cloud rendering approach based on optimized BVH algorithm is proposed.The data points from the WRF(Weather Research and Forecasting) grids are used as cloud primitives,and a Z-order Hilbert curve is employed for spatial sorting.The BVH algorithm based on the Surface Area Heuristic (SAH) is optimized by locally optimizing the cloud primitive density,aiming to enhance computational efficiency.To tackle the data access overhead of overlapping BVH nodes,a novel storage structure called Overlapping Node Sets (ONS) is introduced,which reduces the time complexity.The optimized BVH algorithm reduces unnecessary intersection tests between rays and triangle surfaces,and resolves issues related to invalid boundary volume overlaps.Simulation experiments demonstrate that the proposed method achieves a 15.6% improvement in SAH cost compared to similar state-of-the-art algorithms,a 10% improvement in EPO(Effective Partial Overlap),and a reduction of over 100% in construction time.The computational efficiency of the optimized BVH algorithm outperforms similar algorithms in any WRF cloud scenario,indicating its capability for rapid rendering of WRF cloud products.
CHEN Jiaming , CHEN Xu , REN Shuo , DI Hongwei
2025, 17(2):227-234. DOI: 10.13878/j.cnki.jnuist.20230921003
Abstract:Texture extraction,a pivotal task in computer vision,significantly influences the accuracy of texture classification.Traditional single-texture extraction methods often fail to accurately describe the characteristics of various textures.To address this issue,this paper proposes a texture extraction approach based on an Improved Position Local Binary Pattern (IPLBP) and Gabor filters.The proposed IPLBP enhances texture description capability by integrating texture position information into the LBP framework.Specifically,the IPLBP algorithm captures local texture nuances,while Gabor filters extract global texture attributes.Subsequently,these two complementary feature sets are fused and classified using Support Vector Machine (SVM).Experimental results demonstrate that the proposed approach exhibits excellent performance in texture material classification tasks.Notably,compared to traditional LBP algorithms,the IPLBP-Gabor filter approach more accurately discerns the subtle differences between diverse texture features,thereby enhancing texture classification accuracy.
HUANG Jingjing , WU Wenxuan , TIAN Yu , WANG Can , WANG Maofa
2025, 17(2):235-244. DOI: 10.13878/j.cnki.jnuist.20240617001
Abstract:Here,a bearing fault diagnosis method based on recurrence analysis and Stacking ensemble learning is proposed to effectively extract nonlinear information from rolling bearing signals and improve diagnostic accuracy.Firstly,the nonlinear information in bearing signals is mapped to a two-dimensional recurrence plot through the application of recurrence analysis theory.Convolutional Neural Network (CNN) and Support Vector Machine (SVM) models are established from the perspectives of image recognition and recurrence quantification analysis,respectively.Finally,the Stacking method is employed to integrate these two models,leveraging their respective strengths.Experimental results demonstrate that the proposed method significantly improves the classification accuracy of bearing vibration signals and exhibits excellent stability under varying load conditions,providing a reliable solution for bearing fault diagnosis.
XUE Muhui , XU Baoguo , LI Lang , SONG Aiguo
2025, 17(2):245-255. DOI: 10.13878/j.cnki.jnuist.20240512002
Abstract:In the field of Brain-Computer Interface (BCI),the recognition of natural hand movements through electroencephalography (EEG) is crucial for achieving natural and precise human-machine interaction.However,attempts to enhance model generalization ability across different subjects using transfer learning are still rare in studies focusing on natural hand movement paradigms.Here,we investigate three natural hand movement paradigms of grasping,pinching and twisting through EEG experiments,and validate the effectiveness of two transfer learning algorithms,namely CA-MDM(Covariance matrix centroid Alignment-Minimum Distance to Riemannian Mean) and CA-JDA(Covariance matrix centroid Alignment-Joint Distribution Adaptation),on our experimental dataset.The results show that CA-JDA achieves average accuracies of 60.51%±5.78% and 34.89%±4.42% in binary and quadruple classification tasks,respectively,while CA-MDM performs at 63.88%±4.59% and 35.71%±4.84% in the same tasks,highlighting the advantages of Riemannian space-based classifiers in handling covariance features.This study not only confirms the feasibility of transfer learning in natural hand movement paradigms but also aids in reducing calibration time for BCI systems and implementing natural human-machine interaction strategies.
ZHENG Yan , XI Kuan , BA Wenting , XIAO Yujie , YU Wei
2025, 17(2):256-264. DOI: 10.13878/j.cnki.jnuist.20240506001
Abstract:To address the issues of low search efficiency,long distance,and non-smooth paths in traditional path planning algorithms for autonomous vehicles,this study proposes an improvement by using key nodes of the optimized ant colony algorithm to replace the local target points in the dynamic window approach.Additionally,a target distance evaluation sub-function is incorporated into the dynamic window approach's evaluation function to enhance the efficiency and smoothness of path planning.Furthermore,a path decision-making method is employed to solve the problem of global path failure,enabling the vehicle to avoid obstacles and meet safety requirements of path planning.The improved ant colony algorithm utilizes the positional information of the start and end points to create an uneven initial pheromone distribution,thereby reducing time consumption during the initial search phase.By maintaining the global optimal paths and enhancing the pheromone concentration of excellent local paths,the pheromone update mechanism is optimized to speed up path exploration efficiency.The planned path is further optimized to reduce redundancy in nodes and turning points,thereby shortening path length.Simulation results show that compared to traditional path planning algorithms,the proposed integrated algorithm achieves better performance in terms of distance,smoothness,and convergence,aligning with the safety requirements for autonomous vehicle operation.
2025, 17(2):265-272. DOI: 10.13878/j.cnki.jnuist.20240424002
Abstract:A trajectory planning method based on safety potential field and polynomial lane-changing model is proposed to address the safety and comfort issue of intelligent vehicles in highway lane-changing scenarios.First,the vehicle's motion is decoupled into horizontal and vertical dimensions in the Frenet coordinate system,and the horizontal d-t and the longitudinal s-t trajectory clusters are generated by fifth- and fourth-order polynomials,respectively.Second,to improve algorithm efficiency,a trajectory evaluation index incorporating acceleration,acceleration change rate and curvature is designed according to vehicle dynamics characteristics,and candidate trajectories are obtained after initial screening of the trajectory clusters.Finally,a trajectory evaluation function that integrates safety,comfort and efficiency is established based on safety potential field theory and the concept of minimum driving safety distance,to select the optimal candidate trajectory and complete the simulation verification.The proposed approach is simulated and verified by building a high-speed two-lane curve model and designing diverse lane-changing scenarios for uniform and variable-speed traffic flows.Results show that in the lane-changing process,the collision risk value between the ego-vehicle and obstacle vehicles remains below the critical threshold,ensuring lane-changing safety.Furthermore,under different driving conditions,the ego-vehicle's acceleration,acceleration change rate,and trajectory curvature are all within acceptable limits,indicating that the lane-changing trajectory planning approach ensures both comfort and smoothness in various obstacle traffic flows.
2025, 17(2):273-281. DOI: 10.13878/j.cnki.jnuist.20230523002
Abstract:This paper investigates the event-triggered distributed filtering problem for a class of linear systems with additive and multiplicative noises transmitted over sensor networks,in which the considered process noise and measurement noise exhibit one-step autocorrelation and two-step cross-correlation characteristics.Firstly,a recursive equation is used to describe the dynamic bias of the system,and random variables following a Bernoulli distribution are introduced to characterize the random packet loss phenomenon.Secondly,an event-triggered mechanism is introduced to reduce the information transmission frequency while ensuring filtering performance,and a novel consistency-based distributed filter is constructed.Then,a recursive equation for the upper bound of filtering error covariance is established using stochastic analysis techniques,and an expression for filtering gain is derived by minimizing the variance constraint index.Finally,the effectiveness of the proposed optimized filtering method is verified through numerical simulations.
CAO Haiou , YU Limin , ZHANG Lin , LIU Zhiren , HAN Xiao , DAI Zhihui
2025, 17(2):282-292. DOI: 10.13878/j.cnki.jnuist.20240409002
Abstract:The abnormalities in the host device protection sampling loops of smart substations are difficult to detect due to their characteristics of concealment,transience,and instability.Here,we propose an early warning scheme for the abnormal sampling loops of the host device protection in smart substations based on homologous waveform recording data.First,the occurrence of abnormal sampling loops is determined by reasonably setting the early warning thresholds and waveform analysis criteria.Second,a knowledge graph for the operation and maintenance of the sampling loops is constructed based on information of the equipment operation and maintenance manual related to the sampling loops,to realize the aided decision-making for handling abnormal sampling loops.Finally,a case study is carried out based on an actual defect in a substation.The results show that the scheme is effective in detecting the abnormalities in the sampling loops,and the constructed knowledge graph can provide decision-making references for operation and maintenance personnel in handling abnormalities,significantly reducing manual workload.
FEI Bin , SHEN Haiping , QUE Yunfei , CONG Leyao , JIANG Yiwen
2025, 17(2):293-300. DOI: 10.13878/j.cnki.jnuist.20240505001
Abstract:Considering the spectral characteristics of substation noise,an Enhanced Generative Fixed-Filter Active Noise Control (EGFANC) approach is introduced to address the problems of slow convergence speed,weak tracking capability,and large computational complexity that perplexed adaptive algorithms.A lightweight one-Dimensional Convolutional Neural Network (1D CNN) is employed to output the weight vector based on noise frame information,then the weight vector is combined with sub-control filters to adaptively generate suitable control filters for various types of noise.The simulation results demonstrate that the EGFANC approach has superior noise reduction performance and robustness when dealing with dynamic noise and transformer harmonic noise.In addition,the proposed EGFANC approach can significantly reduce convergence time by selecting appropriate pre-trained control filters for different types of noise.
Address:No. 219, Ningliu Road, Nanjing, Jiangsu Province
Postcode:210044
Phone:025-58731025