Detection of Metal Surface Defects Based on YOLOv4 Algorithm

Abstract

To solve the problem of low recognition accuracy and low defect location accuracy in traditional detection of surface defects of metal materials, this paper innovates on the basis of YOLOv4 architecture, and studies the influence of adding feature pyramid network module to different position of model neck on detection algorithm. Experiments have shown that adding the feature pyramid network (FPN) module after sampling on the neck network can enhance the feature information expression ability of the feature map originally input to the detection head in the size of 80×80 and 40×40, and achieve better detection results, and achieve better detection results. The experimental results show that adding feature pyramid network module to the neck can effectively improve the detection accuracy of the algorithm. Finally, compared with the traditional YOLOv4 network, the average recognition accuracy of this model can reach 92.5% and the recognition accuracy is improved.