Multi-classification recognition and quantitative characterization of surface defects in belt grinding based on YOLOv7

Due to the complex and varied surface defects of the belt grinding, the automatic recognition and quantitative characterization of the surface defects is still an issue that needs to be resolved. To characterize surface defects in belt grinding, this study proposed a multi-classification recognition and quantification method. The YOLOv7 algorithm was utilized to detect defects, which would obtain classification and location information of defects. Aiming at the defects with continuous geometric boundaries, an object detection strategy based on Hough straight-line image correction was designed to quantitatively describe the characteristics of the defects in the grinding direction. In order to capture the quantitative information of burn defects without continuous geometric boundaries, an adaptive threshold segmentation approach with correction factors was developed. The results demonstrated that the YOLOv7 performance on our dataset reaches the mAP of 0.907. The suggested quantitative characterization method can effectively quantitatively characterize and analyze the defects.