Metal Surface Defect Detection Based on Weighted Fusion

摘要

The texture of the metal surface is irregular, and the distinction between defect feature and texture feature is not obvious, which brings great difficulties to automatic defect detection. Traditional methods of defect detection have low accuracy, difficult design, and poor robustness, and the speed and accuracy of existing deep learning methods cannot reach the high standards in the factory. In this paper, a YOLOv4 defect detection algorithm based on weighted fusion is proposed. It uses a GAN network to generate a mask map in real time, and performs weighted fusion of the feature maps in YOLOv4. Experimental results show that the proposed method improves the recognition accuracy by 1% without seriously affecting the speed of the YOLOv4.