CUE-YOLO: An Efficient and Robust Detector for Metal Surface Defect Detection With Adaptive Attention and Task Alignment

Reliable detection of surface defects on metal materials is critical for ensuring product integrity and maintaining high manufacturing standards. However, existing object detection models often struggle with (a) subtle defect representation, (b) classification–localization misalignment, and (c) severe class imbalance, particularly in industrial scenarios. To address these challenges, we propose CUE-YOLO, a task-driven architectural refinement of the YOLOv11 framework, specifically tailored for efficient industrial surface defect inspection. Specifically, the novel C3K2 with Adaptive Sampling Attention (C3K2-ASA) module introduces an attention-guided feature refinement strategy to amplify subtle and fine-grained defect patterns. The designed Unified Task-Aligned Head (UTAH) mitigates task misalignment by harmonizing classification and localization branches through a unified prediction structure. Furthermore, the proposed Exponential Moving Average SlideLoss (EMA-SlideLoss) dynamically adjusts IoU thresholds to provide adaptive supervision, particularly for low-confidence detections and instances from underrepresented classes, thereby improving robustness in the presence of class imbalance. Experiments on the challenging GC10-DET dataset demonstrate that CUE-YOLO achieves a state-of-the-art performance of 74.2% mAP, outperforming the YOLOv11 baseline by 11.9%. Cross-dataset evaluations on aluminum and NEU-DET further verify the generalization ability of the proposed method across diverse industrial domains.