Abstract: In practical production scenarios, the irregular geometric morphology and complex spatial distribution of coal particles not only affect segmentation accuracy but also make manual annotation of segmentation masks extremely inconvenient, limiting their applicability to large-scale industrial scenarios. To address this problem, a dual-stage adaptive segmentation framework (DASeg) for coal particle images was proposed. The framework consisted of the DS-YOLO object detection model, the Adaptive Box Refinement (ABR) module, and the SAM2 image segmentation model. The DS-YOLO model introduced the Dynamic Upsampling (DySample) module and the Spatial and Channel Synergistic Attention (SCSA) module into the neck network of YOLOv11, which effectively improved object detection accuracy. To solve the problem that the detection boxes generated by DS-YOLO did not closely fit the actual coal particle boundaries, the ABR module was designed. The ABR module performed weighted fusion of the original detection boxes and the bounding boxes of the masks according to weighting coefficients to generate more accurate prompt boxes. The corrected coordinate information was then used as prompt input for the SAM2 model, which extracted global and local features and fused prompt region information to generate target masks, thereby achieving coal particle segmentation. Experimental results showed that the DASeg framework performed excellently in coal particle image segmentation tasks, with a Pixel Accuracy (PA) of 93.1%, a Mean Intersection Over Union (mIoU) of 88.4%, and a Mean Dice (mDice) of 93.4%.