Image segmentation for thin structures using a zero-shot learner

Accurately pinpointing thin structures like surface cracks is pivotal for preventive structural maintenance. Given the unpredictable shapes, locations, and dimensions of surface cracks, image segmentation becomes the preferred objective for most research efforts. In contrast to image representation learning, which benefits from well-established image-level representations and a large number of pre-trained models, low-level (pixel-level) image segmentation tasks often necessitate specialized training on specific datasets to achieve accurate and precise results. In this study, we scrutinize the world’s first general-purpose foundation model for image segmentation, the Segment Anything Model (SAM), in the context of thin crack detection. While SAM produces visually satisfactory results, its numerical performance in thin crack evaluation falls short due to discrepancies with the expected ground truth of achieving 1-pixel precision in detection.