Abstract
Crack segmentation is one of the important aspects of infrastructure maintenance and safety. Conventional methods require extensive manual work and are highly prone to human errors, encouraging automated and advanced solutions in this specific field. We propose a new MSMC-U-Net model for improving the precision and reliability of concrete crack segmentation with a systematic three-stage framework. First, data pre-processing is performed to make the data more consistent and less sensitive to variations in the dataset, which involves performing tasks such as resizing images, normalization, and data augmentation operations, which include scaling, rotation, and flipping. Then, the feature extraction module incorporates the concepts of multi-scale and multi-context analysis in conjunction with cutting-edge elements; Transformers learn the sequence context and long-range relationships, convolution block attention module (CBAM)s provides spatial and channel attention to important parts of the input, and feature pyramid pooling (FPP) and atrous spatial pyramid pooling (ASPP)capture multi-scale context information to address different sizes and directions of cracks. In the final stage, the segmentation module employs the U-Net decoder that fuses, upsampled and smooth the output to generate the segmentation masks. The proposed MSMC-U-Net is tested on the CrackSeg9k and CSD datasets and delivers excellent performance with respect to precision, recall, ODS, OIS, mIoU, AP, and FPS scores of 96.87 %, 95.31 %, 95.29 %, 95.36 %, 94.29 %, 96.43 %, and 19 on the CS9k dataset, and 95.44 %, 94.05 %, 93.62 %, 94.75 %, 92.36 %, 95.16 %, and 21 on the CSD dataset, respectively.