CMAC-Net: a cross-modal alignment and semantic consistency optimization framework for image-text fusion in industrial defect detection

Abstract

Industrial defect detection requires accurate identification and localization of fine-grained anomalies, while the inconsistency between visual anomaly responses and textual defect semantics remains insufficiently addressed. To improve the semantic consistency and interpretability of industrial defect detection, this paper proposes CMAC-Net, a Cross-Modal Alignment and Consistency Network for image-text fusion. First, a dual-encoder feature extraction module models local textures, global structures, and defect-related textual semantics from industrial images and descriptions. Then, a hierarchical disentangled cross-modal alignment module separates modality-shared and modality-specific representations and establishes fine-grained correspondences between local defect regions and textual semantics. Furthermore, a text-guided semantic consistency optimization module enhances multi-level visual residual features with aligned textual information, thereby refining visual anomaly responses. Experimental results demonstrate that CMAC-Net achieves competitive performance in industrial defect detection, pixel-level localization, and image-text semantic matching. Specifically, CMAC-Net improves mIoU by 0.83% over Reverse Distillation and AUPRO by 0.92% over SimpleNet on MVTec AD, improves mIoU by 15.82% over CutPaste on VisA, and increases R@1(i2t) by 3.24% over DecAlign on Flickr30k. These results demonstrate that incorporating hierarchical cross-modal alignment and text-guided semantic refinement can effectively strengthen the correspondence between defect regions and textual descriptions, providing a more reliable framework for interpretable and semantically consistent industrial defect detection.