Anti-interference framework for weakly supervised thermal defect segmentation on high-rise building facades.

Infrared thermography is crucial for ensuring the structural safety of high-rise building facades, yet accurate defect quantification typically requires expensive pixel-level fully supervised annotations. While weakly supervised semantic segmentation reduces annotation costs, existing methods struggle with low-quality pseudo-labels and inaccurate boundaries in infrared images. To address this problem, this paper proposes an anti-interference-driven segmentation framework for infrared facade defects, in which the SAM 2 segmentation stage is trained under weak supervision using box prompts and pseudo-mask guidance. Firstly, an improved DeepLabV3+ interference rejection module is constructed, incorporating a boundary-aware triplet loss and simple linear iterative clustering superpixel consistency constraints to eliminate pseudo-thermal anomalies. Secondly, YOLOV11 is trained to localize defects on the interference-suppressed images and provides bounding box prompts for the subsequent SAM 2 segmentation stage. Finally, a weakly supervised fine-tuning strategy based on the segment anything model 2 is proposed, where box prompts and pseudo-mask supervision are used for pixel-level mask learning without relying on manually annotated defect masks during the segmentation training stage. Validated on 180 infrared images from 50 high-rise buildings, the method achieves interference rejection (mean Intersection over Union of 76.2%) and precise localization (mean average precision at 0.5 of 86.5%). Furthermore, it attains pixel-level segmentation with 63.4% Miou, 87.0% F1, 83.1% precision, 91.3% recall, and 92.2% mAP@0.5 using only box-level annotations. This study demonstrates the feasibility of high-precision thermal defect segmentation for large-scale facade inspection.