Steel surface defect detection based on Improved YOLOX Algorithm

Abstract

As one of the very common industrial products, strip steel plays a very important role in various fields. Steel production has received extensive attention. In industrial production, scratches are unavoidable due to numerous forces beyond human control. Cracks and other flaws appear, and their presence has a direct impact on the product’s quality. Therefore, it is of great significance to develop efficient and accurate strip surface defect detection. In this study, the most widely used metal material strip steel is selected as the research object, and its surface defect detection problem is studied accordingly. The activation function and loss function are improved on the basis of the YOLOX model. Resolve the problem of overexposure and content distortion of the image caused by the high reflecting characteristics of the strip, which degrades the image quality, and the embedding of the sub-attention mechanism to enhance the regional characteristics of hidden small targets. The experimental results show that the mAP of the improved model for strip surfaces defect detection is 80.7%, which is 4.1% higher than that of the original YOLOX algorithm, and the improved model can reduce the false detection and missed detection rate of defects.