Enhancing Magnetic Ring Defect Detection With Partially Adaptive Context-Enhanced Module Plugged Into Feature Pyramid Network

Quality control of magnetic rings in complex industrial manufacturing processes is of great importance to ensure optimal performance. Magnetic ring surface defect inspection, being a pivotal aspect of quality control, poses persistent challenges attributed to the unpredictable nature of defect shape and scale, foreground-background imbalance, and other factors. Most current state-of-the-art (SOTA) defect detectors prioritize improving multiscale feature fusion to enhance defect recognition while neglecting the significant influence of improving backbone features. To tackle these challenges, we propose a partially adaptive context-enhanced module (PACEM) that can be seamlessly integrated between the backbone network and the feature pyramid network (FPN) to extract attention-guided contextual features, enabling adaptability to irregular variations and neighboring textures of surface defects. PACEM leverages a partially adaptive context extractor (PACE) to effectively extract and consolidate semantic information across diverse receptive fields, followed by a contextual feature gating module (CFGM) that employs a unique multihead inverted spatial attention (MHISA) mechanism to discern and select influential contextual features. In addition, we present a lightweight decoupled head (LDH) specifically tailored for magnetic ring surface defect detection. Extensive experiments consistently demonstrate that the proposed method exhibits superior performance in enhancing magnetic ring surface defect detection.