Unsupervised Crack Segmentation with Candidate Crack Region Identification and Graph Neural Network Clustering

This paper introduces an innovative approach to unsupervised crack segmentation. Our method initiates with the application of image processing techniques to identify potential crack regions within the images. Leveraging the Canny edge detection algorithm, we delineate edges within the images, followed by the implementation of morphological image processing to eliminate noise. Subsequently, contour analysis is employed to pinpoint candidate crack regions with precision. These identified regions are then input into our unsupervised crack segmentation model, which relies on graph neural network clustering to delineate and categorize the cracks effectively. Experimental results on the CRACK500 dataset [19] showcased the robustness of our approach, evidenced by a Mean Intersection over Union (MIoU) score of 65.88 and a Mean Absolute Error (MAE) of 0.7. Moreover, the proposed method gave the comparable results with the state-of-the-art supervised crack segmentation algorithms.