Research on defect detection method for end face of steel coil based on Faster-RCNN and YOLOv3

At present, there are few researches on the detection of steel coil end face defects, and the existing detection methods cannot take into account the accuracy and speed of the problem. An improved Faster-RCNN and improved YOLOv3 detection algorithms are proposed. Since Faster-RCNN loses the underlying information during continuous sampling, it is difficult to detect small defects on the end face of the steel coil. It is proposed to use multi-scale feature fusion to improve the optimization. At the same time, due to the poor performance of the Faster-RCNN backbone network, the detection rate is slow. It is proposed to use A better-performing residual network alternative. In view of the inaccurate positioning of YOLOv3, it is proposed to use GIoU to replace the IoU in the original YOLOv3. Likewise, to further improve the YOLOv3 detection rate, a residual network is used instead of Darknet-53. Experimental analysis shows that the detection rate of the improved Faster-RCNN_R50 reaches 12.47Fps, the original Faster-RCNN is 8.35Fps, an increase of 4.12Fps, the detection accuracy (AP) of Faster-RCNN_R50 reaches 92.1%, and the original Faster-RCNN is 81.2 %, an increase of 10.9%; the improved YOLOv3_R50 detection rate is as high as 22.87Fps, the original YOLOv3 is 19.07Fps, an increase of 3.80Fps, the YOLOv3_R50 detection accuracy rate reaches 89.5%, the original YOLOv3 detection accuracy rate is 68.7%, an increase of 20.8% . The conclusion shows that in the absence of real-time detection requirements, the improved Faster- RCNN_R50 has the highest accuracy, and if there is a real-time detection requirement, the improved YOLOv3_R50 has the fastest speed, while also ensuring a certain detection accuracy.