Cycle-CFM: An unsupervised framework for robust multimodal anomaly detection in industrial settings

Industrial multimodal anomaly detection is confronted with three pivotal challenges: cross-modal feature drift, noise sensitivity, and modality imbalance. To address these issues, we propose Cycle-Consistent Cross-Modal Feature Mapping (Cycle-CFM), an unsupervised framework that integrates cycle-consistent cross-modal mapping with channel-attention-guided adaptive loss weighting. Cycle-CFM establishes bidirectional feature alignment between RGB and 3D modalities via reversible cycle mappings, yielding consistent representations robust to vibration and depth noise. To further mitigate dynamic interferences such as illumination variations, we introduce a joint optimization strategy that combines cross-consistency and cycle-consistency losses. Experimental results on our self-constructed SteelDefect-3D-AD dataset demonstrate that Cycle-CFM achieves an AUPRO@1 % of 0.371, outperforming state-of-the-art methods by 17–45 %. It also attains a pixel-level AUROC (P-AUROC) of 0.991 and an image-level AUROC (I-AUROC) of 0.998. On the public MVTec 3D-AD benchmark, Cycle-CFM reaches a mean P-AUROC of 0.960 and improves accuracy by 37.5 % for elongated anomalies. With a runtime of 11.03 FPS and 469.52 MB of parameters, the model highlights both its effectiveness and deployability for real-time industrial inspection.