An Improved YOLOv5 Metal Surface Defect Detection Model

Abstract

Surface defect detection is a key step in the process of product inspection. Aiming at the shortcomings of traditional metal surface defect detection in efficiency and accuracy, this paper proposes an improved metal surface defect detection model based on YOLOv5. The improved YOLOv5 model realizes bidirectional cross-scale connection and fast normalized fusion by improving the feature fusion structure and adds a small object detection layer. The performance of the improved model is improved and the average accuracy is 95.3%, which is 2.0% higher than original YOLOv5 model, the number of missed detections is also reduced.