6 Conclusion
⑴ In this paper, the Faster RCNN model and the RetinaNet model are studied. The surface defects of Si3N4 ceramic bearing inner ring are taken as the research object. The backbone of the RetinaNet model is ResNet-50. The performance of RetinaNet model is improved by comparing Faster RCNN with RetinaNet. SE, CBAM and NAM attention modules are added to the tail of ResNet to form attention modules. Comparing the effects of different attentional modules on model performance, NAM attentional module has the most significant effect on model performance. NAM attentional module is used to form improved RetinaNet model.
⑵ The surface defects images are collected by the platform established by ourselves and made into dataset by LabelIMG. The online data augmentation is used to expand the surface defect dataset of Si3N4 ceramic bearing inner ring. The improved Faster RCNN model and RetinaNet model are used to train the surface defect dataset and verify the model performance. This improved RetinaNet has the characteristics of high precision in pit, wear, crack and snowflake detection, and is superior to Faster RCNN model. Compared with the Faster RCNN model and the RetinaNet model, the mAP value of the improved RetinaNet model increased by 13.45% and 2.1%, respectively.