An Improved YOLOv5-based Method for Surface Defect Detection of Steel Plate

Abstract

In the digital manufacturing industry, since various features of steel products cannot quantify through traditional algorithms of quality evaluation, deep learning has been introduced into the field of detecting defects increasingly. However, there are still challenges to be overcome. Most steel plate surface defects are of various types and small, and the detection speed needs to meet the requirements of real-time. In this paper, a new network structure called CI-YOLOv5 be proposed, which is based on YOLOv5 and improves the precision of defect detection while meeting the real-time requirements by improving the network structure of YOLOv5. Two attention mechanisms and involution operator are embedded in CI-YOLOv5 to enhance the ability of feature learning, suppress unimportant information during feature extraction and aggregate richer semantic information. And then the extracted features are directly used to predict the location and classify the defects. Compared with the YOLOv5s, the evaluation metrics of CI-YOLOv5 performance have been improved. Specifically, it has achieved 96% mAP and improved by 5% over YOLOv5s, precision improved by 9% to 94%, recall improved by 1%, and F1 improved by 5% while maintaining the real-time detection of defects.