SFPE-NET: Bearing and Gear Defect Detection Network with Edge Detail Perception and Scale-Residual Feature Fusion

Bearings and gears are prone to surface defects during service, which compromise equipment stability and reliability. To enhance quality control and maintenance efficiency, efficient defect detection methods are urgently needed. However, existing methods have difficulty capturing fine edges and effectively integrating multi-scale features, limiting accurate detection of complex and small defects. To this end, this paper proposes a novel defect detection method, SFPE-NET, which consists of three innovative modules: the Edge and Structure Detail Awareness module (ESDA), the Scalewise Residual Fusion Module (SRFM), and the Reparameterized Detail Enhancement Head (RDEH). Specifically, ESDA includes the Edge Enhancement Module (EEM) and the Shared Feature Pyramid Convolution Module (SFPC). EEM enhances edge representation to improve the perception of complex defects, while SFPC integrates multi-rate convolutions to capture multi-scale features. SRFM adopts partial convolution for efficient multi-scale feature extraction, significantly reducing computational cost while enhancing feature expressiveness. RDEH combines reparameterization with detail-enhancing convolution to optimize the computational path and improve detection accuracy for complex surface patterns. Experiments conducted on a self-built dataset demonstrate that SFPE-NET improves precision, recall, and mean average precision (mAP) by 6.2%, 2.5%, and 4.9%, respectively, while reducing the number of parameters by 15%. The method achieves high performance and efficiency, showing strong potential for defect detection in gears and bearings.