DSCANet: Dynamic Snake Convolution with Attention for Crack Segmentation

Pixel-wise crack segmentation task plays a crucial role in infrastructure maintenance. However, it poses a significant challenge due to the irregular and slender nature of cracks. The standard convolution kernel struggles to accurately capture these structural features of cracks as its shape is a fixed square. Moreover, the presence of intricate details in cracks means that shallow information containing more spatial and nuanced details significantly impacts the segmentation results. Any inadequacy or inaccuracy in local detailed information can lead to segmentation errors. In this paper, we propose a novel crack segmentation method based on encoder-decoder framework. First, we propose a Dynamic Snake Convolution with Attention (DSCA) module to enhance feature extraction accuracy for cracks and direct the network's focus towards critical features. Additionally, we propose a multi-level and multi-scale feature fusion strategy to enable the network to effectively leverage both local spatial information and global semantic information. We enrich the decoder with more detailed information at various levels. Also, we incorporate a channel prior convolutional attention mechanism for feature fusion to supplement attention to both channel and spatial aspects. Finally, a Strip Pooling Module (SPM) is employed to our network. The SPM enable networks to efficiently model long-range dependencies. Experimental results on two different crack datasets validate the superior performance of our method, surpassing several mainstream methods.