YOLO-CSL: a lightweight real-time fabric defect segmentation network for industrial detection

To address the challenges of fabric defect detection and segmentation under complex textured backgrounds, including the fragmentation of continuous defects (e.g.broken yarns and scratches), frequent false and missed detections induced by strong texture interference, inaccurate defect boundary segmentation, and insufficient real-time performance, this paper proposes YOLO-CSL, a lightweight real-time instance segmentation framework based on YOLO11n-seg. The proposed method develops a problem-oriented collaborative optimization framework that progressively enhances defect representation through structure enhancement, texture-aware feature refinement, and lightweight mask reconstruction. Specifically, a C3k2-ConvNeXtDilated module is introduced to enlarge the receptive field and strengthen multi-scale contextual representation of defects. An improved SimAM attention mechanism is incorporated to adaptively recalibrate feature responses by suppressing irrelevant texture interference and enhancing defect-related information. In addition, a Lightweight Multi-scale Context Attention (LMCA) segmentation head is designed to aggregate multi-scale boundary features while reducing computational complexity. Experimental results show that the proposed model achieves mAP50 scores of 0.873 and 0.870 on detection and segmentation tasks, respectively, outperforming several mainstream YOLO-based instance segmentation models. These results demonstrate that YOLO-CSL provides an effective and efficient solution for real-time fabric defect detection and segmentation under complex textured backgrounds, showing strong potential for industrial online detection.