<p>Coal-rock interface recognition is crucial for intelligent and green coal mining, with significant implications for mining safety and resource utilization. However, in underground coal mines, coal-rock images captured under low-illumination and dusty conditions often suffer from reduced contrast, blurred boundaries, and weakened texture details. Existing methods struggle to achieve a satisfactory balance between recognition accuracy and computational efficiency. To address these issues, this study proposes a degradation-aware and boundary-constrained coal-rock interface recognition network termed CRFormer. Specifically, a degradation-aware feature enhancement module is designed in the first encoder stage to adaptively compensate degraded shallow features and enhance the representation of texture and structure around the coal-rock interface. A stage-wise hybrid attention mechanism assigns cross-shaped window attention to low-level features and full-window attention to high-level features, achieving global context modeling while maintaining computational efficiency, which is crucial for practical on-site deployment. In addition, a boundary-constrained auxiliary loss is introduced to complement the cross-entropy loss, explicitly improving coal-rock interface continuity and localization accuracy. Experimental results on a self-built coal-rock dataset show that the model achieves a pixel accuracy (PA) of 98.97% and a mean Intersection over Union (mIoU) of 97.92%. Compared with several representative image segmentation networks and recent coal-rock interface recognition models, CRFormer exhibits a better balance between model recognition accuracy and inference speed.</p>