Research on defect detection algorithm of strip steel based on improved YOLOv4

Abstract

To address the current problems of wide range of strip steel surface defect size variation, slow detection efficiency, low detection accuracy, and difficulty of mobile-side model deployment, an improved YOLOv4 algorithm model is proposed in this paper. Firstly, in order to improve the robustness of the model, data augmentation is applied to the dataset. Secondly, in order to improve the matching between the a priori frame and the feature map, the K-means++ algorithm with faster convergence and better results is used instead of the K-means algorithm in the original YOLO algorithm for the design of the a priori frame. Finally, CSPDarknet is specifically replaced for the Ghostnet to enhance the backbone network's ability to extract defective features. The experimental results show that the improved YOLOv4 algorithm achieves 87.9% mAP on the publicly available NEU-DET dataset, which is 2.4% lower than the original YOLOv4 algorithm. However, the number of parameters of the model decreases by 80% compared with the original YOLOv4, and the detection speed is around 44 FPS, which can not only meet the needs of industrial production, but also meet the requirements of deploying the model to mobile.