CIFormer: Cross Interaction TransFormer for surface defect detection in ceramic tile

Surface defect detection in ceramic tiles plays an important role in industrial intelligent manufacturing. However, detecting ceramic surface defects remains challenging due to: (1) minute and intricate defect characteristics; and (2) visual similarity between defects and unaffected regions. To address these challenges, a Cross Interaction Transformer (CIFormer) is proposed for surface defect detection in ceramic tiles, where the Coarse-Fine Granularity Interaction (CFGI) module is designed to capture intricate defect regions by modeling both coarse and fine-level features, and the Adaptive Cross-Scale Attention (ACSA) module is developed to differentiate the similar defects from the normal regions by aggregating information across multiple scales. Experiments on the benchmark validate the effectiveness of the proposed CIFormer, which outperforms state-of-the-art methods with performance scores of 77.5% Precision, 72.7% Recall, and 77.5% mAP50. Specifically, it surpasses the best existing method (CTDD-YOLO) by 2.3% in Precision, 2.5% in Recall, 2.8% in mAP50, and 3.6% in mAP50-95, demonstrating comprehensive improvements across all key metrics and further enhancing the quality of ceramic tiles in industrial manufacturing.