The problem the researchers set out to solve is geometric rather than exotic. Fabric defects such as holes, neps, knots and oil stains often occupy a vanishingly small fraction of the image, while coarse ends, weft shrinkage, broken ends and density bars stretch along the warp or weft direction with a very narrow short side. In the Tianchi dataset, 72.3% of annotated boxes occupy less than 1% of the original image area, and the median relative area is just 0.096%. The periodic weave of the fabric itself further suppresses the grayscale and texture contrast between weak defects and the background, so a detector can easily merge tiny anomalies into the normal texture response.

The scale statistics become especially telling after aspect-ratio-preserving letterbox resizing to the 640 by 640 pixel input canvas that the detector uses. The original images are 2446 by 1000 pixels, so the letterbox scale factor is about 0.2617, and the median defect shrinks to an area-equivalent side of only 12.66 pixels, with a median true short side of 6.7 pixels. On the highest-resolution detection layer of standard YOLOv11n, called P3/8 because each grid cell corresponds to an 8 by 8 pixel region of the input, the median defect spans only about 1.58 grid intervals. Measured along its true short side, the coverage drops below one cell. For an anchor-free detection head, such a coarse grid leaves very few spatial locations inside or near a tiny box, which can undermine the assignment of positive training samples.

FD-YOLO addresses this by extending the detection pyramid from three levels to four. The authors retained the YOLOv11n backbone, its FPN-PAN feature-fusion neck and its decoupled detection head, but continued the top-down upsampling path one more level and fused the result with the shallow backbone P2 feature. This creates a P2/4 detection branch with a stride of 4, halving the finest sampling stride and quadrupling the global prediction-grid density relative to P3. Under a 640 by 640 input, the P3 layer offers 6,400 prediction locations while P2 offers 25,600. In terms of geometric coverage, the median defect now spans about 3.16 grid intervals on P2/4, roughly double its P3 coverage, and a representative 20-pixel box covers 5 by 5 intervals on P2 instead of 2.5 by 2.5 on P3.

The authors were careful to separate geometric sampling density from the actual assignment of training samples, which is governed by YOLOv11’s TaskAlignedAssigner. This assigner scores candidate predictions by combining classification confidence with overlap quality, selects the top candidates for each ground-truth box, and resolves conflicts. Instrumented measurements on the 1,183-image validation set showed that for 1,290 input-space small targets, adding the P2 branch increased the mean number of eligible centers per ground truth from 7.59 to 13.37 and the final positive samples from 4.87 to 6.89, while mean positive-sample IoU rose from 0.498 to 0.586. Notably, these per-target increases are far smaller than the fourfold global grid-density ratio, and P2 received 88.5% of the positive assignments for small targets, confirming that the new layer genuinely absorbs the tiny-defect workload.

Controlled ablations support the conclusion that the denser detection grid, rather than some incidental capacity increase, drives the improvement. An alternative variant that replaced the backbone’s downsampling convolutions with an information-preserving ADown operator reached only 45.5% mAP@0.5, and combining ADown with the P2 branch yielded 47.4%, well below the 52.6% of the P2-only model. Variants using attention modules, alternative upsampling operators such as DySample, or regression losses such as Wise-IoU and Shape-IoU topped out at 49.2%. To rule out brute-force scaling, the team also trained a widened three-scale YOLOv11n with 4.166 million parameters and 10.315 GFLOPs, matched to FD-YOLO’s compute budget; averaged over three seeds it reached only 44.86% mAP@0.5, below even the baseline, while FD-YOLO achieved 51.23% under the same budget.

Error analysis at a fixed confidence threshold of 0.25 shows the gain is not bought with more false alarms. The baseline produced 809 true positives, 432 false positives and 1,112 false negatives, whereas FD-YOLO produced 894 true positives, 413 false positives and 1,027 false negatives, recovering 85 additional defects while making 19 fewer false detections. Scale-stratified evaluation showed AP@0.5 improving in all three COCO size bins, from 46.26% to 52.15% for small objects, 53.93% to 57.59% for medium and 42.55% to 46.58% for large. Feature visualizations with EigenCAM reinforced the mechanistic picture: baseline activations spread across broad texture regions, while FD-YOLO’s stride-4 layer produced finer hotspots closer to the actual defect locations.

The advantage also holds when category labels are kept and when the data change. On the complete 20-class Tianchi task, mAP@0.5 rose from 28.4% to 31.5% and recall from 31.4% to 36.0%, with AP improving for 15 of 20 classes; the largest gains came for pattern skip, broken spandex, three-filament, knot and hole, though five classes declined, reflecting persistent long-tail and inter-class confusion. On the independently trained AITEX dataset under a balanced-patch protocol at 512 input resolution, FD-YOLO reached 56.32% mAP@0.5 versus 49.36% for the baseline, although the authors caution that the test set contains only 28 positive patches, giving wide confidence intervals. FD-YOLO also outperformed protocol-matched DETR, Deformable DETR and RT-DETR baselines, which scored between 40.8% and 48.5%, while using 2.66 million parameters against their 20.0 to 41.2 million.

Resolution sensitivity provides further evidence for the stride-coverage mechanism. When models trained at 640 pixels were evaluated at inference resolutions from 416 to 1024 without retraining, FD-YOLO exceeded the baseline at all five resolutions, with the gap widening to 10.1 percentage points at 416 pixels, where the median defect covers only about 1.03 P3 grid intervals, and narrowing as resolution, and hence P3 coverage, increases. Both models peaked at 800 pixels. The authors note this trend reflects inference-resolution sensitivity rather than a controlled causal effect, since training occurred only at 640.

The costs and remaining limits are stated plainly. The P2 branch raises computation from 6.30 to 10.20 GFLOPs because it operates on a higher-resolution feature map, and measured batch-one FP32 latency on an RTX 4090 D rose from 7.55 to 8.58 milliseconds per image. Even with the improvements, roughly half of defects are still missed in the full 20-class task, and extremely elongated targets can be missed or localized with low IoU, with some texture-induced false detections remaining. The authors therefore position FD-YOLO as better suited to a first-stage screening role that flags suspicious regions for human verification rather than fully autonomous accept-or-reject decisions, and they note that the experiments are confined to a single train-validation split per dataset, leaving boundary localization, category discrimination and large-scale cross-fabric validation as open challenges.

Subject of Research: A high-resolution P2/4 detection branch added to YOLOv11n to improve tiny fabric defect detection

Article Title: Tiny fabric defect detection with a high resolution branch in YOLOv11n

Article References: Sun, Y., & Liang, Z. (2026). Tiny fabric defect detection with a high resolution branch in YOLOv11n. Discover Artificial Intelligence, 6(1), Article 1574. https://doi.org/10.1007/s44163-026-02553-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02553-w

Keywords: fabric defect detection, YOLOv11n, FD-YOLO, small object detection, P2 high-resolution branch, computer vision, textile quality control, Tianchi dataset, AITEX, positive-sample assignment, object detection, deep learning