Surface Defect Detection of Aluminum Material based on Deep Learning

Abstract

As one of the important materials required for industrial production, aluminum profiles are being widely used in industry. However, in the process of producing aluminum profiles, different kinds of defects will appear on the surface of aluminum profiles, so timely and effective detection of their surface defects is particularly important for later manufacturing. At present, most factories still focus on manual visual inspection, which requires a lot of labor costs and the inspection standards are not uniform. Therefore, it is particularly necessary to achieve high-efficiency and rapid detection of aluminum surface defects.Since deep learning can learn independently and has strong adaptability, target detection algorithms based on deep learning have gradually begun to be applied to defect detection. In this paper, three mainstream target detection algorithms, Faster R-CNN, SSD and Yolov5, are used to detect surface defects on aluminum, and the experimental results show that the average accuracy (mAP) of Faster R-CNN can reach up to 79.49%, and the speed of Yolov5 (fps) can reach 49. The experimental results prove that Faster R-CNN is effective in target detection accuracy, while SSD and Yolov5 are more effective in terms of detection of target rate.