Crack identification and severity analysis via computer vision: Comparative study of rgb and grayscale imagery

Deep learning-based computer vision technology has become an effective tool for crack detection and segmentation, and it has considerable potential for application in infrastructure maintenance. Crack detection and segmentation are the two core steps in automated crack identification. The present study used RetinaNet to detect concrete cracks in RGB and grayscale images and then used three deep learning-based semantic segmentation models to segment the detected cracks: an edge-detection-based bidirectional cascade network (BDCN), a holistically nested edge detection (HED), and a multitask road extractor (MTRD) with a multitask learning framework. The results indicated that RetinaNet exhibited considerably higher crack detection performance on grayscale images than RGB images. In particular, it exhibited a precision of 1.0 under an intersection-over-union threshold of 0.1-0.3. Moreover, regression analysis revealed that the coefficient of determination (R2) for the correlation between crack direction and severity was 0.830, indicating that a crack's direction significantly affected its severity. The adopted segmentation models also performed better on grayscale images than on RGB images, with the accuracy rates of the MTRD and BDCN models for grayscale images being 0.996 and 0.982, respectively. Thus, grayscale processing can enhance crack identification. Practically, this approach can significantly improve the efficiency and reliability of infrastructure inspection and maintenance, enabling quicker identification and prioritization of critical structural repairs.