CrackVAN: A detection method of crack hazard images based on multi-scale feature refinement and large kernel attention

Detection and segmentation of mine cracks are a crucial step in the process of eliminating the safety hazards of mine caused by large-scale mining on the floor of mines. However, the effectiveness of the current segmentation network still needs to be improved due to various reasons, especially the interference of shadows and the edge details of the cracks. In order to enhance the efficacy of crack detection, a multi-scale dilated larger kernel attention model has been developed. Furthermore, utilizing the visual attention network (VAN) architecture, we propose a multiscale dilated visual attention network (MD-VAN) for the early-stage screening of images containing cracks. In the crack detection stage, we propose a segmentation network, CrackVAN, which is based on multiscale feature refinement using MD-VAN as an encoder. In the feature processing part, we propose a multiscale feature refinement module (MSRF) for processing deep features, which enhances the information of cracks in the deep features by multiscale DWConv. Additionally, a convolutional pyramid compressed attention (CPSA) is proposed to further optimize the shallow features, enhancing the crack boundary information by increasing the attention at both the spatial and channel levels. Experimental results on the mine dataset show that MD-VAN achieves accuracy (94.12%) compared to other networks, while CrackVAN achieves mIoU (92.13%) and F1-score (95.76%). In addition, an evaluation of the efficacy of the proposed methodology was conducted on two publicly available datasets, namely CFD and DeepCrack, and our method gets better performance.