Tire quality is vital for vehicle safety, but finding surface defects automatically is hard. Current systems often fail to spot tiny flaws hidden inside complex tread patterns. Also, because defects vary in size and blend into the dark rubber, standard models usually output broken or messy segmentation masks. To solve this, we built YOLO11-SPMB, a new instance segmentation network based on YOLO11. We designed it specifically for tire inspection. First, we added a Multi-Scale Adaptive Feature Fusion (MS-AFF) module to the network. It filters out the repeating noise from normal treads so tiny defects can stand out. Second, we created a Boundary Enhancement and Imbalance Calibration (BE-IC) strategy. It uses a boundary Dice loss to keep the mask edges smooth and whole. At the same time, it adjusts class weights to help the model catch rare but major damage, fixing the data imbalance problem. Tests on our custom dataset show the model works very well. YOLO11-SPMB achieves a mAP@0.5:0.95 of 0.702, a mIoU of 0.782, and a recall of 0.928. It also runs at 60 frames per second (FPS), making it fast and accurate enough for real factory production lines.