Abstract: To solve the problems of low image segmentation efficiency, missed detection of small targets, and poor boundary segmentation accuracy caused by overlapping boundaries in images with dense rock debris distribution, a fine-grained segmentation network for rock debris images was developed. This model was based on the Mask R-CNN network architecture. It used DW lightweight convolution to build the backbone network and utilized coordinate attention mechanism to construct a multi-scale feature pyramid to enhance feature selectivity and capture spatial position information, so as to improve the detection ability of targets of different sizes. By combining the distance graph-guided boundary enhancement loss function, the model’s learning of fine-grained boundary features was further optimized, and the accuracy of segmentation results was improved. The improved rock debris segmentation network was tested using rock debris image samples collected from blocks such as Fenggu and Xinchang. The average precision was 92.1%, and the mean intersection over union(mIOU) was 83.2%. The results show that the improved rock debris segmentation network has significantly improved inference speed and accuracy compared to existing segmentation networks, providing technical support for automated rock debris image analysis.