A Frequency-Aware Multi-Stage Framework for Accurate Magnetic Tile Surface Defect Inspection

Surface defect detection in magnetic tiles is essential for automated quality inspection. However, spatial sparsity of defect regions, pervasive background noise, and low surface contrast lead to weak feature representation, deep feature oversmoothing, and noise amplification under global enhancement. To address these challenges, we propose a three-stage framework integrating localization, enhancement, and segmentation. In the localization stage, we design a Cross-Stage Partial Peripheral Frequency Gating Network (CSP-PFGNet) backbone and an Attention-based Intra-scale Feature Interaction with Wavelet Directional Attention Module (AIFI-WDAM) layer, leveraging frequency-guided gating and directional attention to suppress background clutter and refine multi-scale defect representation. Subsequently, Locally Adaptive Contrast Enhancement (LACE) is applied exclusively within defect regions of interest to amplify contrast while isolating noise, followed by Co-Representation Purification (CoRP) to recover pixel-level defect masks via cross-image commonality mining. Experimental results show that the proposed framework improves mAP50 by 1.76 percentage points while reducing parameters and GFLOPs by 11.07% and 11.40%, respectively. In the final segmentation stage, it achieves an IoU of 84.15% and a pixel accuracy of 99.34%, significantly outperforming state-of-the-art baselines.