WECI-YOLO: Multi-Module Enhanced Steel Surface Quality Defect Detection

Steel surface quality is crucial for industrial product performance, yet defect detection remains challenging due to type diversity, scale variations, and complex backgrounds. Existing deep learning methods often lack accuracy and robustness. This paper proposes WECI-YOLO, an enhanced algorithm based on YOLO11n. The backbone network integrates a C3k2-WTConv module to widen the receptive field. A lightweight EUCB module is embedded in the upsampling path to improve multi-level feature fusion. For elongated defects, a C2PSA-CAA attention module is designed to optimize extraction. Additionally, the Inner-SIoU loss function combines SIoU’s angle perception with Inner-IoU’s internal scaling strategy. Experiments on the GC10-DET and Aug-GC10-DET datasets demonstrate that WECI-YOLO achieves recall rates of 65.8% and 66.5%, respectively, along with mean average precision (mAP) values of 66.6% and 67.9%. Compared to YOLO11n, recall improved by 2.4% and 2.5%, while precision increased by 4.0% and 3.6%, achieving optimal performance on Aug-GC10-DET. The impacts of augmentation and the proposed modules are evaluated separately in the experiments.Notably, these accuracy gains are achieved with only a small and acceptable increase in computational cost. The proposed algorithm demonstrates clear advantages over existing methods in steel surface defect detection.