Enhancing pixel-level crack segmentation with visual mamba and convolutional networks

Computer vision-based semantic segmentation methods are currently the most widely used for automated detection of structural cracks in buildings and pavements. However, these methods face persistent challenges in detecting fine cracks with small widths and in distinguishing cracks from background stains. This paper addresses these issues by introducing MambaCrackNet, a new network architecture for pixel-level crack segmentation. MambaCrackNet incorporates residual visual Mamba blocks and integrates visual Mamba and convolutional neural network-based segmentation techniques. This approach effectively enhances the detection of fine cracks, reduces misdetections of background stains, and remains robust to variations in patch size and training sample sizes, making it highly practical for engineering applications. On two open access crack datasets, MambaCrackNet outperformed mainstream crack segmentation models, achieving MIoU scores of 0.8939 and 0.8560 and F1-scores of 0.8817 and 0.8412.