Abstract: The difficulty in accurately quantifying fracture features at underground excavation faces constrains stope blastability evaluation and refined blasting operations. An intelligent workflow—fracture recognition, indicator quantification, blastability classification, and scheme matching—was developed by integrating a You Only Look Once (YOLO)-based segmentation model with a comprehensive cloud model. Multimodal data augmentation was designed to emulate harsh underground imaging conditions, expanding 119 labeled images to 474 for model training. With DeepLabV3 and YOLOv8l-seg as reference models, YOLOv5-seg variants were benchmarked using Dice, pixel accuracy (PA), mask intersection over union (Mask IoU), and box intersection over union (Box IoU). YOLOv5x-seg was then selected and further enhanced by replacing PANet with a bidirectional feature pyramid network (BiFPN) and embedding a convolutional block attention module (CBAM), yielding YOLOv5x-CBF with consistent metric gains. Fracture trace length was quantified within the effective face region and integrated with powder factor and blasting advance. These indicators were then used to construct a comprehensive cloud model for uncertainty representation, three-class blastability classification, and “class–scheme” mapping. Field-scale trials on two stopes demonstrated improved blasting performance, with blasting advance increased by 18%–30% and powder factor reduced by 8%–13%. These results confirm the practicality of the proposed method and provide a field-deployable technical reference for intelligent blasting in underground metal mines.