Surface defect detection of stainless steel pipes based on YOLO-DSDS model

Abstract Stainless steel pipes, as the main material for oil and chemical transportation systems, have essential surface quality, and the surface quality of steel pipes after pickling-cleaning is usually inspected manually. With the acceleration of production pace, improvement of product quality, and manpower shortages, manual inspection methods can no longer meet current demands. Therefore, a vision-based method for detecting surface defects on stainless steel pipes is proposed. First, a detection device for the spiral conveyance of steel pipes was designed through simulation. Then, a YOLO-DSDS object detection model was proposed, which incorporates a depth aware convolution (DAC) module into the Backbone network to improve feature extraction capability. During the feature extraction process, a downsampling module was designed that simultaneously adds Shuffle Attention (SA) and Spatial Pyramid Pooling Fast (SPPF) to enhance the contextual information in the feature maps. Finally, the Wise-IoU (WIoU) loss function was used to calculate the loss value, improving the overall performance of the detector. Experimental results show that this detection method achieved a mean Average Precision (mAP) of 90.37% on a self-made dataset, improving by 5.58% and 1.77% compared to YOLOv7 and YOLOv9 respectively. In the publicly available NEU-DET and GC10-DE datasets, the model achieved mAP values of 80.03% and 77.84%, respectively, which is superior to other models.