DAC-Net: deep attention and feature fusion U-net for bearing surface defect detection under complex conditions

Bearings are indispensable components in industrial production, but detecting their surface defects is challenging due to complex backgrounds. To address this, we propose DAC-Net, a high-precision model for bearing surface defect detection. DAC-Net integrates advanced deep attention mechanisms and a novel feature extraction architecture. The MEC-PA module improves global feature capture and edge preservation, while a MaxPooling-Dropout hybrid enhances generalization and reduces overfitting. The DBLE module enables multi-scale fusion, boosting segmentation precision. The MUFF and CBAMT modules refine feature utilization through spatial and channel attention, improving target detection and noise suppression. Additionally, the PACA module extracts key image details, and the PCAP module integrates pixel- and channel-level features for fine-grained segmentation. Extensive experiments on a publicly available bearing defect dataset show that DAC-Net outperforms existing methods, achieving Recall increases by 4.55%, Precision by 2.19%, IoU by 4.87%, F1 score by 3.97%, and MCC by 3.72%.