Research On Steel Surface Defect Detection Based On YOLOv5s

The number of surface defects is always the key evaluation criterion in steel production.Traditional detection methods,represented by the strobe method,suffer from drawbacks such as slow detection response and low accuracy.Object detection technology based on deep learning can effectively improve this problem due to its strong real-time performance and high accuracy.In actual production,industrial factors such as complex production environments can affect the performance of object detection technology,resulting in decreased detection efficiency.Therefore,improving the detection speed and accuracy of surface defects on steel has significant research implications.This paper focuses on steel surface defect detection tasks,using the YOLOv5 s object detection algorithm and industrial steel,the most common material in industrial environments,as the research object.The main purpose is to improve the problems of slow detection speed and low accuracy in steel surface defect detection tasks.The main work is as follows:To address the problem of inaccurate target positioning in steel defect detection tasks,we use K-means++ clustering Anchor and add detection layers to improve the model’s ability to locate defect targets.To address the difficulty of extracting defect features,we add a CA attention mechanism to the YOLOv5 s backbone network and introduce the window self-attention mechanism in the Swin Transformer to enhance the model’s feature extraction ability.At the same time,adding the CA mechanism and the window self-attention mechanism to the feature fusion network can respectively weaken background information interference and reduce the loss of information of small targets in defects,which is beneficial to enhance the feature fusion ability of the network.Due to the limitations of the CIOU loss function calculation method in YOLOv5 s,replacing it with the SIOU loss function can enhance the robustness of the network.To address the problem of relatively dark backgrounds in the NEU-DET steel defect dataset,we use contrast enhancement and brightness enhancement for image preprocessing to highlight defect features.The experimental results show that the proposed model has strong steel surface defect detection capabilities,with an m AP value of 79.5%,which is 6.2% higher than the original YOLOv5 s m AP value of 73.3%,and the missed detection and false detection cases have been improved.Compared with current mainstream algorithms SSD and Faster R-CNN,the improved model still has advantages in accuracy.Furthermore,the proposed model was lightened,and the experimental results showed that the m AP value of the lightened network was 78.6%,which is 0.9% lower than the non-lightened model,and the detection speed increased from 71.9FPS to 80.1FPS.The problem of slow speed and low precision in steel surface defect detection task is improved.This provides a certain reference value for the practical application of steel surface defect detection tasks.