Ms-unet:coal-rock crack image segmentation based on attention mechanism and multi-scale features
摘要
The accurate segmentation of coal rock cracks is crucial for the coal mining industry. The use of existing mainstream segmentation methods often leads to the fragmentation and loss of overall cracks, as well as the occurrence of roughness and false positives in local details. Based on this, this paper designs a Channel Attention Atrus Space Pyramid Pooling Module (TD-ASPP) and Cross Attention Non Local Blocks (CANB), and integrates them into TransUNet to propose a Coal Rock Crack Image Segmentation Network (MS-UNet) based on attention mechanism and multi-scale features. MS-UNet accurately achieves crack segmentation while eliminating false positives in the segmentation mask.We conducted extensive experiments on the CrackForest dataset and coal rock crack dataset, and compared them with current advanced mainstream means, both achieving SOTA results(CrackForest dataset: