Abstract
Infrared cable anomaly detection in complex scenes is challenging because anomaly samples are scarce, backgrounds are cluttered, and thermal anomalies are often subtle and localized. This paper proposes a mask conditioned diffusion reconstruction method for infrared cable anomaly detection. The target cable-region mask is introduced as spatial prior information to guide normal-pattern learning, while truncated diffusion reconstruction is used during inference to preserve structural information and improve anomaly highlighting. Image-level anomaly scores are further constructed from masked residual maps. Experiments on synthesized infrared cable anomaly data demonstrate the effectiveness of the proposed method, achieving an AUC of 0.9855 under the optimal setting. Ablation studies validate the contributions of mask conditioning, truncated-step selection, and residual scoring.