loading page

Recognition and classification for surface defects of Si 3 N 4 ceramic bearing inner ring based on RetinaNet method and NAM attention mechanism
  • +3
  • Nanxing WU,
  • Dahai LIAO,
  • Zhihui CUI,
  • Xin ZHANG,
  • Wenjie LI,
  • Qi ZHENG
Nanxing WU
Laboratory of Ceramic Material Processing Technology Engineering Jiangxi 333403 PR China)

Corresponding Author:[email protected]

Author Profile
Dahai LIAO
Laboratory of Ceramic Material Processing Technology Engineering Jiangxi 333403 PR China)
Author Profile
Zhihui CUI
Laboratory of Ceramic Material Processing Technology Engineering Jiangxi 333403 PR China)
Author Profile
Xin ZHANG
Laboratory of Ceramic Material Processing Technology Engineering Jiangxi 333403 PR China)
Author Profile
Wenjie LI
Laboratory of Ceramic Material Processing Technology Engineering Jiangxi 333403 PR China)
Author Profile
Qi ZHENG
Jingdezhen Ceramic Institute
Author Profile

Abstract

Due to surface defects on inner ring of Si 3N 4 ceramic bearing are tiny and difficult to defect, the defects accelerate the wear of ceramic parts and reduce the performance of ceramic parts. A surface defect detection method based on RetinaNet method and NAM attention mechanism is proposed. Besides, the performance of RetinaNet method and Faster RCNN method is compared. The platform for surface defects of Si 3N 4 ceramic bearing inner ring is built independently to collect images. The dataset is made up of collected images and expanded by online data augmentation. Resnet-50 is used as the feature extraction network. The NAM attention mechanism is added to the tail of Resnet-50 to form an attention module to improve the model accuracy. As the bounding box regression loss function, loss is used for learning bounding box regression and localization uncertainty. A multi-scale feature pyramid is constructed by a feature pyramid network to integrate multi-level feature information. And a small full convolutional network is used as a classification sub-network and a bounding box regression sub-network. The results show that the mAP of the method reaches 91.84%, which is 13.45% and 2.1% higher compared to Faster RCNN and RetinaNet, respectively. The method has good detection effect on the identification and classification of surface defect species.