STC-YOLO: small target defect detection method for surface texture of Si ₃ N ₄ ceramic balls

Si3N4 ceramic ball in ceramic bearings is a key component for the stability of aerospace equipment, and its surface scratches, pits, wear, spalling, and snowflake defects may cause accidents, so it is necessary to detect its quality. However, the existing detection methods mostly use top-down layer-by-layer extraction for feature learning and ignore the texture and location of small defects in multi-scale, resulting in insufficient learning of small features, which leads to defect leakage detection. In order to improve the detection ability of multi-type defects and the generalisation of multi-type defects detection, a small target detection method for multi-type ceramic ball surface texture defects (STC-YOLO) is proposed in this paper. Firstly, a shallow feature extraction module was designed to retain the semantics of the shallow feature maps and be used for feature fusion. Secondly, a deep and shallow parallel feature fusion module was designed, focusing on both local and global features, to address the issue of losing the local details and position information of small targets during the feature fusion process. The method was evaluated on the SOD-SNCB dataset, and the mean average precision (mAP50) of small target defect detection reached 0.992. The results show that STC-YOLO has excellent detection performance.