Abstract
Coal fracture segmentation in CT images is critical for coal structure analysis, coalbed methane extraction, and mine safety, but it is challenged by complex fracture features and limited computing resources for mine-site deployment. Basic UNet exhibits redundancy, sensitivity to image noise, and high overfitting risk. This study proposes LightECA-UNet, integrating depthwise separable convolution (DSC), efficient channel attention (ECA), and adaptive channel pruning. Experiments show LightECA-UNet achieves 1.6% higher mean Intersection over Union (mIoU) and 2.5% higher fracture IoU than currently popular models. Compared to lightweight counterparts, it reduces computational load by 87.1% and parameter count by 86.9%, enabling deployment on mine-used edge equipment while maintaining segmentation accuracy.