Research on Iron Surface Crack Detection Algorithm Based on Improved YOLOv4 Network

Abstract

Metal surface defect detection has always been an important branch of target detection. In certain places, if the cracks on the metal surface can be found in time, the existing safety hazards can be eliminated. In this paper, a polarized imaging camera with strong environmental applicability is used for sampling, and the degree of polarization image is applied to the detection of iron material cracks. The iron material polarization image crack data set (PICD-iron) is established, and the Cascade-YOLOv4 (C-YOLOv4) network model is used for crack detection. Then it solves the problems of target detection in the dark environment, complex and variable crack detection, small target detection. Experimental tests verify that the detection accuracy of the C-YOLOv4 net-work has improved while comparing with the YOLOv4 network, the detection speed has also increased by 28%.