Automatic identification and early warning of surface cracks have become urgent for mine safety. The precise segmentation of mine crack images is a challenging task due to the complex nature of their content. The segmentation precision can be hindered by the presence of complex environmental backgrounds, such as crack sizes, illumination, slope, and vegetation, which can all interfere with the accuracy of the detection process. In this paper, a deep learning based mine crack segmentation method is proposed, which includes data preprocessing, crack image and non-crack image classification, and crack segmentation. The segmentation network is different from the existing DeeplabV3 Plus, Segformer, etc. It extricates three distinct scales of features within the encoder and proposes a unique multi-feature fusion and attention approach to fuse these three features in the encoder. We have also constructed a unique mine crack dataset. Experimental results on this dataset show that our method obtains 83.48% mIoU and 90.34% F1-score, demonstrating higher segmentation accuracy than other state-of-the-art methods.