A YOLOv5s-GC-based surface defect detection method of strip steel

The surface defect detection of strip steel is very important for identifying the steel quality which has a great significance for improving the production efficiency and ensuring the products’ quality. However, the traditional strip surface defect detection algorithm cannot effectively balance the detection speed and detection accuracy that leads to a low practical detection efficiency. Therefore, a novel detection method based on YOLOv5s-GC has been proposed. Firstly, the ResNet-Mini has been developed to pre-classify the original dataset to reduce the amount of operation of the YOLOv5s-GC network. Secondly, an image preprocessing approach has been proposed to enhance the defect features, which contains two steps. The first step is combining the ResNet-Mini network weights with Grad-CAM to crop the defective areas, and simultaneously removing the background interference; the second step is applying the OTSU and Normal Distribution Enhancement Algorithm (NDEA) to extract the feature grayscale. Thirdly, size enhance strategy has been introduced to obtain a larger size of the data which can simulate the actucal application scenario with large area saccade of strip steel. Finally, the CBL module of YOLOv5s has been replaced by GhostBottleneck and the convolutional block attention module (CBAM) has been added in the Neck layer. From the final experimental results, relative to the original YOLOv5s algorithm, the YOLOv5s-GC’s average accuracy in the detection of 6 types of strip steel has reached to 77.5%, which is 7.9% higher than YOLOv5s, and detection accuracy rate of the average precision is 95.6%, at the same time, the calculation amount of YOLOv5s-GC is reduced by 48% and the inference speed is increased by 24%.