GeoFusionNet: A geometry-aware fusion network for bearing surface defect detection

Abstract

As critical rotating components, bearings are susceptible to surface defects with diverse morphologies and complex textures, posing challenges for defect detection and pixel-level localization. Existing methods neglect the geometric priors of bearings and lack bearing-specific noise synthesis strategies. To address these issues, we propose a geometry-aware fusion network (GeoFusionNet). Focusing on the annular structure of bearings, a segmentation-detection-fusion pipeline is constructed. A lightweight semantic segmentation network first partitions the bearing images, followed by polar coordinate transformation and independent defect detection within each functional region. Finally, global result fusion is achieved through inverse transformation. In the detection phase, we design a simple anomaly detection network integrated with two enhancement mechanisms. Realistic noise synthesis generates structurally plausible pseudo-defects in feature space to improve detection performance. Feature channel selection adaptively retains discriminative channels to suppress redundancy and reduce model complexity. To meet the demands of industrial defect inspection, we introduce a framework for model calibration and maintenance. GeoFusionNet attains a 67.27% pixel-level average precision on the bearing defect dataset, outperforming other mainstream methods. Experiments demonstrate the effectiveness of the proposed method and its favorable trade-off between inference efficiency and detection performance.