CrackSAM: Study on Few-Shot Segmentation of Mural Crack Images Using SAM

Due to the scarcity of mural disease data and the complexity of mural backgrounds, Traditional deep learning image segmentation requires a large amount of annotated data for training, which limits its segmentation performance. Although the SAM (Segment Anything Model) excels in natural image segmentation with its zero-shot segmentation capabilities and geometric prompts, it performs poorly in segmenting mural cracks and requires manual input prompts during the inference stage. This study proposes an novel mural crack segmentation method, CrackSAM. First, the high-performance image encoder from the SAM model is used to generate image embedding feature maps of the input images. Next, our designed feature extraction module further refines these features, enhancing the model's ability to capture crack details. Then, a decoder module upsamples and decodes the refined features to generate highresolution segmentation masks. This process ensures the accuracy of the segmentation results. In experiments, even with only a small amount of public crack datasets for training, our method demonstrates performance close to specialized crack segmentation networks trained on full datasets. Notably, on our self-constructed mural crack dataset, the method effectively distinguishes and segments cracks in complex backgrounds, proving its potential and effectiveness in practical applications. This achievement provides a new technical approach for the automated segmentation and assessment of mural damages.