A New Method of Surface Defect Semantic Segmentation of Steel Ball Based on Pre-Trained U-Net Model

The steel ball surface defect is one of the main reasons for affecting bearing performance. It is very necessary to recognize the defects from steel ball surface images. Sematic segmentation is a vital step in recognition. Therefore, a new method for steel ball surface defect segmentation is presented. The method which used pre-trained U-Net model based on deep transfer learning and could extract objective features. The results showed that the values of MPA and MIoU are 0.7533 and 0.8431, respectively. It demonstrates that the presented method is better than others.