Dilated convolution and edge fusion network for crack edge detection

Abstract Efficient and accurate extraction of pavement crack edges is a crucial prerequisite for the identification of structural road defects. However, existing deep learning methods primarily focus on object detection and segmentation, which makes it challenging to accurately capture the fine-grained edge information of cracks. To address this issue, we propose a dilated convolution and edge fusion network (DCEF) for crack edge detection. The proposed DCEF incorporates a dilated context-aware module to expand the receptive field without increasing the number of parameters, allowing the simultaneous capture of local details and global context. A dual attention module is integrated to highlight crack-relevant features and suppress background noise by modeling spatial and channel-wise dependencies. Furthermore, we design an edge-aware fusion module based on depthwise separable convolutions to efficiently integrate multiscale features while significantly reducing computational complexity. Experimental results in a dedicated crack dataset demonstrate that our method achieves optimal dataset scale and optimal image scale scores of 0.821 and 0.905, respectively, outperforming several state-of-the-art methods. These results validate the effectiveness and practicality of the proposed approach.