Manual crack inspection is labor-intensive and impractical at scale, prompting a shift toward AI-based segmentation methods. We present a novel crack segmentation model that leverages the Segment Anything Model 2 (SAM 2) through transfer learning to detect cracks on masonry surfaces. Unlike prior approaches that rely on encoders pretrained for image classification, we fine-tune SAM 2, originally trained for segmentation tasks, by freezing its Hiera encoder and FPN neck, while adapting its prompt encoder, LoRA matrices, and mask decoder for the crack segmentation task. No prompt input is used during training to avoid detection overhead. Our aim is to increase robustness to noise and enhance generalizability across different surface types. This work demonstrates the potential of foundational segmentation models in enabling more reliable and field-ready AI-based crack detection tools.
