Tiny fabric defect detection with a high resolution branch in YOLOv11n

Fabric surface defect detection is a fundamental task in textile quality control. Fabric defects are frequently embedded in periodic textures and may simultaneously be low-contrast, tiny, and elongated; after repeated downsampling, local anomaly cues are easily weakened before they reach the detection head. Based on the post-letterbox defect-scale statistics, FD-YOLO extends YOLOv11n with a P2/4 high-resolution detection branch to provide more candidate locations for tiny and elongated defects. On the Tianchi 2019 Guangdong Fabric Defect dataset, 72.3% of annotated boxes occupy less than 1% of the original image area. After letterbox resizing to 640× 640 , the median area-equivalent defect side is 12.7 pixels, spanning about 1.6 grid intervals on P3/8 and 3.2 on P2/4. In the merged single-class setting, FD-YOLO attains 52.6% mAP@0.5, a 5.8 percentage-point improvement over YOLOv11n; in the complete 20-class task, mAP@0.5 improves from 28.4 to 31.5% and recall from 31.4 to 36.0%. At 416 inference resolution the single-class gain widens to 10.1 percentage points. These results are consistent with an association between finer detection stride, candidate assignment, and tiny-defect detection, while the improvement also transfers to an independently trained AITEX evaluation. FD-YOLO increases computation from 6.30 to 10.20 GFLOPs, primarily because of the high-resolution feature map.