摘要
针对复杂环境下钢轨表面缺陷检测中由于特征提取不足引发检测精度不良问题,提出一种融合可分离核注意力机制与部分卷积的钢轨表面缺陷检测算法YOLOv11-LSPS。首先,利用大可分离核注意力机制实现卷积核的分解,避免出现因深度卷积导致计算复杂度随核大小增加而二次增长的情况,提高运算效率。其次,在骨干网络中引入Swin Transformer机制以提高对缺陷局部特征的提取能力,增强对小目标特征的耦合性。最后,在颈部网络中引入部分卷积,在减少相关冗余参数计算的同时深度挖掘空间特征,增强模型鲁棒性及泛化能力。采用NEU-DET数据集进行实验验证的结果表明,所提算法相较于传统算法,检测精度提高6.3%,检测速度提升35%。此外,通过在GC10-DET数据集上的验证,证实了该算法的泛化性,即具备一定的工程实践价值。
[Objective]The traditional methods used for railroad surface defect detection are inefficient and exhibit difficulties in identifying subtle or tiny defects,which are critical for ensuring railway safety and product quality.The existing deep learning models exhibit limitations such as inadequate small-target feature extraction,weak multiscale feature coupling,and poor generalization in complex environments.To overcome these limitations,in this study,we combine an attention mechanism(the Swin-Transformer module)with a lightweight convolutional strategy based on the postgraduate course“Condition Detection and Fault Diagnosis.‟The proposed improved algorithm(YOLOv11-LSPS)provides a robust solution for railway safety monitoring.[Methods]We propose an improved YOLOv11 algorithm for railroad defect detection.Our key innovations include 1)C2PSA-LSKA Module:it combines the large separable kernel attention mechanism with C2PSA by decomposing convolutional kernels to improve small-defect feature extraction while reducing computational load;2)Swin Transformer:It replaces the standard C3k2 layer by exploiting the shifted-window self-attention mechanism to improve multiscale feature fusion and long-range dependency learning;3)P-C3k2 Module:It implements partial convolution on selected channels to eliminate redundant computations.The model is trained on NEU-DET and validated on GC10-DET to evaluate its generalization.Metrics(mAP,FPS,Params,and FLOPs)are used as benchmarks for YOLOv11,Faster R-CNN,and YOLOv8-FD and combined with ablation studies to isolate the individual module contributions.[Results]The YOLOv11-LSPS algorithm achieves:1)80.3%mAP on NEU-DET and a 6.3%improvement over the baseline YOLOv11,with notably reduced small-target miss rates;79.8%mAP on GC10-DET,demonstrating strong cross-scene generalization;2)A 35%speed increase(168.7 FPS),meeting real-time industrial inspection standards;3)A compact architecture(5.6 M parameters,8.3 GFLOPs),outperforming the efficiency of two-stage(e.g.,Faster R-CNN)and some one-stage detectors;4)Grad-CAM visualizations,verifying the superior feature extraction performance of the proposed method regarding defects in small targets and complex backgrounds.[Conclusions]By incorporating LSKA and the Swin Transformer,the adaptability of the proposed model to tiny defects and complex backgrounds is significantly improved.Additionally,the computational efficiency of the P-C3k2 module is significantly improved by performing partial convolution,thus balancing performance and cost.In summary,the proposed method achieves an optimal tradeoff among accuracy,speed,and generalization,providing significant practical value in industrial applications.
作者
李书莲
高正中
张经龙
尤振环
LI Shulian;GAO Zhengzhong;ZHANG Jinglong;YOU Zhenhuan(College of Electrical Engineering and Automation,Shandong University of Science and Technology,Qingdao 266590,China;Wenshang Yiqiao Coal Mine Limited Liability Company,Jining 272501,China;Huangdao Customs,Qingdao 266555,China)
出处
《实验技术与管理》
北大核心
2025年第9期111-119,共9页
Experimental Technology and Management
基金
国家自然科学基金项目(62273215)。
