YOLO11-DSU: An enhanced YOLO11 framework for defect detection in dark woven fabrics

Although computer vision technologies have progressed rapidly, leading detectors experience severe accuracy degradation on low-contrast dark fabrics, frequently yielding mean average precision (mAP) scores below 50%. To address these inherent limitations, this work introduces YOLO11-DSU, an enhanced object identification paradigm constructed upon the foundational YOLO11 baseline. To counteract representation degradation encountered within poorly illuminated environments, a novel dual-backbone structure is integrated into the architecture. Furthermore, a space-channel synergistic attention (SCSA) mechanism is embedded at the fusion node, forming an end-to-end optimization process regulated by the adaptive unified intersection over union (UIoU) loss function. This configuration guarantees the deep integration of multilevel features while effectively mitigating noise. Empirical evaluations indicate that the [email protected] of YOLO11-DSU reaches 80.3% on the TILDA dataset and 48.3% on the Tianchi dataset, outperforming the baseline by absolute margins of 6.8 and 4.1 percentage points, respectively. The proposed work establishes a robust technical framework for automated quality testing in high-end textile manufacturing.