BiContext: Multilevel Contextual Enhancement for Efficient and Precise Detection of Magnetic Ring Surface Defects

The surface defect detection task is an integral component of quality monitoring within the manufacturing process of magnetic ring workpieces. In such an industrial scenario, the background tends to be relatively uniform, while the unpredictable defects scattered sporadically across products are intricate and challenging to detect. The significant foreground-background imbalance may result in scarce, but crucial defect features being overwhelmed by background information in deep convolutional operations. To address this problem, we have developed a bilateral context block, integrated into the YOLOv8 detection framework, which establishes two levels of context, progressive local context for efficiently extracting inherent defect features and sparse global context for flexibly capturing neighborhood characteristics to aid in defect recognition. To further emphasize the crucial defect features, we propose a critical spatial feature enhancement module with a straightforward selection mechanism for amplifying foreground information. Furthermore, an interclass gated attention regulator is employed at the end of the BiContext block to adjust the learned defect feature representations adaptively. Extensive experiments conducted on a magnetic ring surface defect dataset, as well as on four publicly available benchmark defect datasets (NEU-DET, GC10-DET, PCB, and Tianchi-Fabric), demonstrate both quantitatively and qualitatively that the proposed method surpasses many competitive detectors and is more lightweight.