Research on Detecting Bearing-Cover Defects Based on Improved YOLOv3

Detecting defects, which is a branch of target detection in the field computer vision, widely used factory production. To solve problems existing algorithms that relate to their insensitivity large or medium defect targets on bearing covers, difficulty detecting subtle defects effectively and lack real-time detection, this work, we establish large-scale bearing-cover dataset propose an improved YOLOv3 network model. The proposed model divided into four submodels: bottleneck attention (BNA-Net), prediction subnet model, localization large-size output feature branch. test generality, robustness practicability new design comparative experiment under abnormal illumination conditions. We ablation verify validity submodules. experimental results show our solves problem algorithm's satisfies mAP result 69.74%, 16.31%, 13.4%, 13%, 10.9%, 7.2% more than YOLOv3, EfficientDet-D2, YOLOv5, YOLOv4, PP-YOLO, respectively.