Abstract: To address the challenges of strong adhesion,high inter-object similarity,and poor image quality in open-pit mine blast pile orerock images,conventional segmentation methods often yield low accuracy and unsatisfactory results. To this end,we introduce an attention mechanism into the traditional U-Net framework and propose an attention optimized CAU-Net method for segmenting blast pile ore-rock images. The proposed architecture retains U-Net as the backbone and incorporates attention modules before each convolutional learning stage. Comparative segmentation experiments are conducted against three attention models. The results show that the CAU-Net model significantly improves segmentation performance,achieving an accuracy(ACC) of 94. 08% and an AUC(Area Under Curve) of 97. 43%,indicating strong predictive capability. Further comparisons with threshold-based segmentation and the standard U-Net,together with evaluation using the Intersection over Union(IoU) metric,confirm that the proposed method can segment blast pile ore-rock images more accurately. The high-precision segmentation results,when applied to digital ore-rock fragmentation statistics,substantially enhance statistical accuracy and effectively advance the digitalization of mining operations.