XB-RTDETR: bearing defect detection based on multi-scale edge feature fusion and intra-scale feature interaction

Bearing surface defect detection is essential in mechanical engineering and industrial inspection because bearings are widely used as core components in mechanical equipment. However, repeated feature extraction and downsampling can remove fine details, making tiny or subtle defects difficult to identify accurately. To address this problem, an XB-RTDETR is proposed for bearing-ring surface defect detection. Inspired by GhostNet, reparameterized convolution (RepConv) is integrated into the backbone to strengthen feature extraction and gradient propagation while reducing computational cost and parameter count. A polarity-aware linear attention mechanism is then incorporated into the Attention-based Intra-scale Feature Interaction (AIFI) module of the Real-Time Detection Transformer (RT-DETR). By retaining informative negative query-key correlations, this mechanism captures more diverse feature interactions. A Local Feature Downsampling Module (LFDM) is also designed to preserve fine details and feature granularity in deep feature maps during downsampling. In addition, a Multi-Scale Edge Feature Fusion (MEFF) module fuses edge-enhanced features with multi-scale backbone representations to improve the detection of subtle edge defects. A bearing surface defect dataset was constructed from images collected at industrial production sites, and the proposed model was evaluated through comparative and ablation experiments. Relative to the RT-DETR baseline, recall (R) increased by 1.4 percentage points, mean average precision at an intersection-over-union threshold of 0.50 (mAP50) increased by 1.0 percentage point, the parameter count decreased by 4.26 million, and the inference speed increased by 6.7 frames per second (FPS). These results indicate that XB-RTDETR improves both detection performance and inference efficiency and is suitable for online bearing surface inspection.