IR-ADS: Invariant Representation and Anomaly Separation for Robust Building Surface Defect Detection

Surface defect detection is crucial for ensuring the structural integrity of buildings. Traditional methods often struggle with distribution shifts caused by unknown materials or defect types. To address this, we propose IR-ADS, a method based on invariant representation learning and anomaly distribution separation. By introducing a depth-enhanced fusion strategy, IR-ADS aggregates geometric cues with RGB features, promoting view-invariant attributes. Additionally, an anomaly-separation mechanism, grounded in Schrödinger Bridge Theory, effectively segments unseen defect types. Extensive experiments demonstrate IR-ADS's robustness and lightweight model complexity, outperforming baselines by over 3.9% in mIoU.