GDS-YOLO11: a multi-scale defect detection method for printed fabrics

Printed fabrics are widely used, but their complex manufacturing process often introduce defects that compromise quality and value, making defect detection essential. Current machine-vision methods struggle to accurately detect multi-scale and irregular defects in these fabrics. To address this, we propose GDS-YOLO11, a multi-scale defect detection approach. To address irregular shapes and strong background interference, we design a C3k2-DCNv4 module with deformable convolution to enhance irregular-defect feature extraction and suppress background noise. To cope with large scale differences, we integrate a GFPN structure into the neck to deepen multi-scale fusion and strengthen feature representation across defect sizes. To improve the detection of defects at different scales, Slide Loss is adopted to adaptively reweight samples based on prediction difficulty. Ablation and comparative experiments show that our method improves mAP50 by 8.0% and recall by 4.5% over the baseline, while also outperforming mainstream detectors in speed and model compactness. This approach provides a more reliable solution for multi-scale defect detection in printed fabrics and offers practical support for intelligent textile inspection.