基于两阶段视觉检测的动车螺栓缺陷识别算法OA

To address the need for defect identification in precision components of electric multiple unit(EMU)bogies during operational image inspection,this paper proposes a recognition algorithm for de-tecting loose and missing bogie bolts by combining template matching with an improved YOLOv5 model.The proposed algorithm first performs bogie extraction.During bogie extraction,bilinear inter-polation is utilized to downscale the images,significantly reducing computational overhead.Simultane-ously,normalized cross-correlation template matching is employed to maintain high-precision target localization on the downscaled images.Then,the proposed algorithm performs defect recognition,for which an improved YOLOv5 model is applied.For the defect recognition task,the model improve-ments mainly include three aspects:first,to address the large parameter size and high computational cost of the baseline model,a CNN-based Cross-Scale Feature-Fusion Module(CCFM)replaces the Path Aggregation Network(PANet),effectively reducing both the parameter count and overall model size;second,to meet the requirements for detecting tiny targets within the dataset,a small-object detec-tion layer is added to enhance the capture of minute defects;third,to overcome the limited feature per-ception range,Large Selective Kernel(LSK)modules are embedded into the shallow network layers to expand the receptive field and improve detection accuracy.Finally,the bogie extraction efficiency and de-fect recognition performance of the algorithm are experimentally verified.Experimental results demon-strate that,in the bogie extraction stage,the average processing time of template matching on down-scaled images is reduced by approximately 92％,from 929.58 ms to 74.87 ms per image.In the defect recognition stage,compared with the baseline model's mean Average Precision(mAP)values of 89.1％at an IoU threshold of 0.5(mAP@0.5)and 43.1％over IoU thresholds from 0.5 to 0.95(mAP@0.5:0.95),respectively,the improved model achieves 91.5％and 46.2％while simultaneously reducing the parameter count by approximately 23％,yielding improvements of 2.4％and 3.1％.The proposed method effectively enhances both the efficiency and accuracy of defect detection for precision components in EMUs,providing a valuable technical reference for safe operation and maintenance.