A nondestructive recognition and classification method for detecting surface defects of Si3N4 bearing balls based on an optimized convolutional neural network

Abstract

Traditional machine vision detection methods have low efficiency and poor detection accuracy in the detection and recognition of surface defects of Si3N4 bearing ceramic balls. In the paper, a nondestructive recognition and classification method of surface defects is proposed. It is based on the improved YOLOv5 algorithm. A novel attention mechanism for mobile networks, Coordinate Attention, is applied to the backbone of the YOLOv5 algorithm to accurately locate and identify defective regions on the surface of Si3N4 ceramic bearing balls. In order to obtain a more effective multi-scale fusion method, fusing the defect feature information at different scales as a weighted bidirectional feature pyramid network, the BiFPN structure is fused in the neck of the YOLOv5 algorithm. The experimental results show that the proposed method achieves a performance with 98.8% mAP and 0.92 average F1-score for surface defect detection and classification. In addition, the detection speed of a signal defect image is 9.9 ms/img.