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

The core challenge in automated fabric inspection is one of separation. A photograph of fabric is essentially a superposition of two things: the regular, low-rank structure of the weave, which repeats predictably across the image, and whatever deviates from that structure, which may be a defect or may simply be noise, lighting variation, or natural irregularity in the material. Robust Principal Component Analysis, a mathematical framework introduced by Emmanuel Candes and colleagues in 2009, formalizes this decomposition by splitting a data matrix into a low-rank component and a sparse component. In the fabric context, the low-rank part corresponds to the background texture and the sparse part to candidate defects. The trouble is that real fabric textures are not perfectly low-rank, and real defects are not perfectly sparse, so the decomposition can blur the boundary between the two, especially when defects are faint.

Zhang and colleagues extend this framework from flat matrices to tensors, the higher-dimensional generalization that allows an image to be represented along multiple simultaneous axes. Their first move is to construct a multi-directional feature tensor using a bank of Gabor filters, a classical computer vision tool that has been used in textured-material inspection since at least the early 2000s. Gabor filters respond selectively to oriented, periodic structures at particular scales, which makes them well suited to capturing the directional threads of a weave. By filtering the fabric image in multiple orientations and frequencies and stacking the responses into a tensor, the method gains a rich, multi-view description of the texture in which genuine defects stand out more clearly than they would in the raw grayscale image.

The second and arguably most novel ingredient is a spatial prior-guided weighted tensor L1-norm. In standard formulations, the sparse component is penalized uniformly, which implicitly assumes that every pixel is equally likely to harbor an anomaly. The new method abandons that assumption. Using kernel density estimation, a non-parametric statistical technique for estimating the probability distribution underlying a set of observations, the researchers generate a prior weight tensor that encodes how anomalous each spatial location appears before the main optimization even begins. Locations whose feature statistics fall in the tails of the estimated distribution receive higher weights, flagging them as more likely defect candidates. This weighting is computed up front and then held fixed during optimization, a design choice the authors describe as achieving numerical decoupling between weight assignment and iterative optimization.

That decoupling matters for a subtle but important reason. If the weights were allowed to change freely during the iterative solution of the decomposition, the optimizer could reinforce its own early mistakes, drifting into a local optimum in which background noise is misclassified as signal. By fixing the prior weights before the alternating iterations begin, AAW-TRPCA prevents the model from being captured by background noise, keeping the decomposition anchored to the statistically informed prior. The approach echoes ideas from outlier detection with kernel density functions and from visual saliency models based on self-resemblance, but integrates them directly into the algebraic machinery of tensor decomposition rather than applying them as a separate preprocessing step.

The third component addresses a different failure mode: not all feature channels are equally informative. Some Gabor responses may be dominated by noise or may fail to discriminate defects for a particular fabric pattern. The researchers introduce an adaptive group sparsity norm driven by structural quality feedback. This mechanism combines two established image-quality measures: the Hoyer sparsity measure, which quantifies how concentrated a signal is in a few large components, and total variation, a spatial compactness measure rooted in the classic 1992 work of Rudin, Osher, and Fatemi on noise removal. Together these form what the authors call a soft veto mechanism, evaluated dynamically during the ADMM optimization process, the alternating direction method of multipliers that solves the decomposition by breaking it into simpler subproblems.

The soft veto works by scoring each feature channel on how well its sparse output satisfies the structural expectations of a genuine defect, namely sparsity in magnitude and compactness in space. Channels that score poorly are progressively down-weighted or filtered out, so that noisy or uninformative directions cannot contaminate the final defect map. Because the evaluation happens inside the optimization loop rather than after it, the method can adapt to the specific characteristics of each fabric image, retaining channels that prove useful and discarding those that do not. This adaptive pruning distinguishes the method from earlier weighted tensor RPCA formulations, such as double auto-weighted approaches, which assign weights based on global statistics rather than on structural quality feedback measured during the solve.

The experimental evaluation, conducted on both simulated data and real-world industrial datasets, including the publicly available TILDA textile defect dataset, shows that AAW-TRPCA suppresses complex background noise while preserving the structural integrity of genuine defects. The most striking result concerns the ultra-low false positive rate regime, the operating point that matters most in a factory. In industrial inspection, a false alarm is not free: every flagged roll must be re-examined by a human or diverted for manual review, so a detector that cries wolf becomes economically useless even if it catches most true defects. The authors report that compared with both traditional methods and deep learning algorithms, the proposed method achieves superior detection performance precisely in this critical regime, where the tolerance for spurious alarms is vanishingly small.

This finding carries broader significance for the field of industrial anomaly detection. Deep learning approaches such as PatchCore, PaDiM, and SimpleNet have dominated recent benchmarks in surface inspection, learning powerful feature representations from large corpora of defect-free images. Yet they typically require substantial training data, careful tuning, and computational resources that may not suit every production line. AAW-TRPCA, by contrast, is unsupervised and model-based: it requires no labeled defects and no learned network weights, relying instead on the mathematical structure of the tensor decomposition and statistically grounded priors. For textile mills producing constantly changing patterns, where collecting defect-free training images for every new design is impractical, such a method offers a compelling alternative that generalizes across patterns without retraining.

The work, funded by the National Natural Science Foundation of China and the Natural Science Foundation of Zhejiang Province, sits at a productive intersection of linear algebra, computer vision, and textile engineering. Its lineage is visible in the references: from Gabor-filter-based defect detection in textured materials, through low-rank recovery methods for patterned fabrics, to recent total-variation-regularized tensor RPCA variants for high-dimensional data recovery. What the Zhejiang team adds is a coherent story about where prior knowledge should enter the decomposition and how quality feedback should shape it, turning a generic separation algorithm into one tuned to the physics of fabric and the economics of inspection. As automated quality control spreads through textile manufacturing, methods that can find the faintest flaw while keeping false alarms near zero may prove decisive, and this study offers a mathematically principled step in that direction.

Subject of Research: Unsupervised fabric defect detection using anomaly-aware weighted tensor robust principal component analysis

Article Title: Fabric defect detection via Anomaly-Aware Adaptive Weighted Tensor Robust Principal Component Analysis

Article References: Zhang, Y., Lu, Y., & Lv, L. (2026). Fabric defect detection via Anomaly-Aware Adaptive Weighted Tensor Robust Principal Component Analysis. Cluster Computing, 29(14), Article 828. https://doi.org/10.1007/s10586-026-06485-z

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06485-z

Keywords: fabric defect detection, tensor robust principal component analysis, unsupervised anomaly detection, Gabor filters, kernel density estimation, total variation, Hoyer sparsity, ADMM optimization, textile quality control, computer vision, low-rank decomposition, industrial inspection