A multi-prompt transfer learning framework based on segment anything model 2 for segmenting microcracks in engineered cementitious composite images and videos

Accurate segmentation of engineered cementitious composites (ECC) microcracks is important for evaluating material damage and cracking behavior. However, ECC cracks are typically slender, low-contrast, and densely distributed, and can easily be confused with pores, shadows, and surface textures. Existing crack segmentation methods rely heavily on large-scale pixel-level annotations and often exhibit limited robustness when applied to ECC microcracks. To address these issues, this study proposes a multi-prompt transfer learning framework based on Segment Anything Model 2 (SAM2). The framework integrates a threshold-based coarse-mask prior, a self-perceptive prompt box, and Euclidean distance transform-guided positive and negative point prompts to enhance crack localization while suppressing noise and preserving crack connectivity. In addition, a progressive unfreezing-based few-shot transfer strategy and a hybrid loss function are employed to effectively adapt the foundation model to the ECC crack domain while reducing annotation costs. Experimental results on a laboratory ECC dataset show that the proposed framework achieves a mean accuracy of 99.1%, a mean intersection over union (IoU) of 86.4%, and a Dice coefficient of 92.6%, outperforming mainstream deep learning baselines. Notably, with only five annotated training images, the framework still achieves a mean IoU of 81.7% and a Dice coefficient of 89.7%, demonstrating strong performance under a highly limited annotation budget. Without any video-specific training, the framework also achieves temporally consistent crack video segmentation through the streaming memory mechanism of SAM2. Overall, the proposed framework provides an effective and annotation-efficient method for detecting microcracks in both static images and dynamic videos.