Modern single-stage and two-stage neural detectors exceed 90 percent accuracy when kernel geometry matches fabric texture

Fabric is not an unstructured natural image. Plain weave and twill contain repeated spatial structure, while many defects interrupt that structure locally. The established taxonomy separates statistical, structural, spectral and model-based methods from learning-based methods. Statistical descriptors include grey-level co-occurrence matrices and autocorrelation. Spectral methods include Fourier, Gabor and wavelet transforms. These approaches encode prior knowledge about cloth rather than asking a network to discover every regularity from examples.[S32]

A Fourier representation converts periodic warp and weft structure into concentrated frequency components. Removing or attenuating the components associated with normal repetition can make a local phase disruption easier to see. A wavelet decomposition retains localisation across scales, which is useful when a fault occupies only a small image region. Gabor filters add orientation and frequency selectivity, so a bank can respond differently to a warp-aligned streak, a weft-aligned streak and the background weave. The survey evidence supports these as established fabric-inspection methods, not as obsolete precursors to neural networks.[S32]

That prior structure has three practical uses. First, it can suppress a predictable background before a detector sees the image, reducing the variation that the model must explain. Second, it provides a diagnostic channel: a change in the dominant spectrum can reveal a new weave, tension change or acquisition fault before defect labels are examined. Third, it offers a compact baseline against which a more expensive model must demonstrate an improvement. These are engineering recommendations derived from the mechanisms documented in the review, not claims of a measured GFT result.[S32]

Classical filters also have clear failure conditions. A fixed frequency mask assumes a stable period and alignment. Stretch, camera skew, a new denting pattern or a change from plain weave to a complex print can move the background energy that the mask was designed to remove. Excessive suppression can erase a faint defect together with the weave. The preprocessing stage must therefore be calibrated by fabric family and preserved under version control alongside the model weights. It cannot be treated as a neutral, universal cleaning step.[S16][S32]

DF-YOLO provides a direct example of the hybrid approach. Its backbone adds frequency-domain perception to spatial information through Fourier transform, while its neck combines an improved bidirectional feature pyramid with convolutional block attention. The stated purpose is better generalisation across fabric textures and anomalies, addressing the limited receptive field of conventional convolution. The important lesson is not that every inspection system should copy DF-YOLO. It is that global periodic context and local spatial evidence answer different questions and can be combined deliberately.[S16]

Gabor processing can play a similar role ahead of learned features. In one-class fabric inspection, a Gabor filter bank supplies texture responses to an autoencoder, followed by density estimation for localisation. That particular model belongs to the cold-start discussion in the next chapter, but its architecture matters here: handcrafted directional responses and learned representations are compatible. A supervised detector can use the same principle without adopting the one-class decision rule.[S06]

GFT's implementation rule is to keep the raw calibrated frame, the preprocessing configuration and the transformed input linked by the same image identifier. An ablation must compare raw input, classical preprocessing alone and the combined model on identical roll-level splits. If the hybrid does not improve critical-defect recall enough to justify its computation and calibration burden, the simpler pipeline wins. This prevents an attractive architectural diagram from becoming an unquestioned production dependency.

Two-stage detectors separate candidate generation from refinement. Faster R-CNN proposes regions and then classifies and localises them. Mask R-CNN adds a pixel mask for each proposed instance. The broader detection literature describes the resulting trade-off: the additional stage can improve localisation and classification precision, but increases computational work and can slow inference. Fabric inspection makes that trade-off specific. A bounding box may be enough to stop a roll, while a mask may be necessary to measure an irregular stain or tear.[S01][S32]

The Mask R-CNN framework reported for fabric inspection uses a ResNeXt-101 backbone, Feature Pyramid Networks and RoI Align. Its reported mAP lies between 93 and 97 percent, with recall near 97 percent and F1-scores between 95 and 96 percent. These are strong reported results for that study's framework and evaluation. They should not be read as a guaranteed mask quality on every weave, nor compared numerically with image-level accuracy from another paper.[S33]

Each component has a distinct job. The backbone extracts hierarchical features. The feature pyramid retains routes for defects of different sizes. RoI Align samples features without the quantisation error associated with coarse region rounding. The mask head then predicts which pixels belong to the candidate defect. Together, these mechanisms produce a boundary that can be converted into physical length and area after the camera calibration has been applied. The source supports the architecture and reported scores; the measurement use follows from the mask output and the optical calibration boundary.[S33]

Pixel-precise output creates a data obligation. Bounding boxes can be annotated relatively quickly, whereas reliable masks require careful treatment of fuzzy edges, partial transmission and regions where human graders themselves disagree. A model trained on inconsistent masks will reproduce that inconsistency with greater confidence, not resolve it. For grading, the annotation guide must state whether a stain boundary is drawn at the visible colour change, at a defined contrast threshold or around the complete affected area. This is a protocol requirement rather than a claim that one drawing convention fits every buyer.

The acceptance metrics must also follow the output. Image accuracy cannot certify a mask. The evaluation should report detection recall, localisation overlap and the error in derived physical dimensions, each on a held-out set. A model can find nearly every defect while systematically drawing a boundary too tightly, which would bias length-based grading. Conversely, a generous mask can preserve recall while overstating severity. Boundary error must therefore be reviewed beside detection error before the mask controls an automated disposition.

Two-stage segmentation is best reserved for decisions that use its additional information. It can run as a second pass on frames escalated by a faster screen, or on a station where cycle time allows detailed measurement. It should not be imposed on every frame merely because its published score is high. The next chapter's cold-start models and the later edge chapter address different parts of this deployment problem; the architectural principle here is simpler: pay for a mask when the physical decision needs a mask.[S01][S33]

YOLO treats detection as a unified prediction rather than a proposal followed by a separate refinement. Its backbone extracts features, its neck fuses information across scales and its head predicts classes and locations. The family has evolved through deeper residual backbones, cross-stage partial connections, feature pyramids, path aggregation, anchor-free prediction and attention mechanisms. The recent review covers this progression from YOLOv1 to YOLOv11 specifically for fabric inspection, including the continuing tension among localisation, classification and processing speed.[S01]

A version number is not a selection criterion. The useful question is which mechanism solves the observed failure. Multi-scale prediction matters when a pinhole and a long reed mark share a frame. Cross-stage connections matter when the compute budget constrains feature reuse. Attention matters when relevant defect responses are weak against a busy background. Anchor-free prediction changes how locations are assigned, but does not remove the need to represent extreme aspect ratios. The architecture should be read as a set of testable mechanisms rather than a chronological ladder in which every later model must be better on local cloth.[S01][S11]

Single-stage detection suits roll screening because the same network can produce a class, confidence and location without a separate proposal stage. That direct path reduces pipeline complexity and gives the edge system a clear output for a roll map. It does not guarantee sufficient recall for small or occluded objects. The review explicitly notes that the efficiency of single-stage designs can come at the cost of accuracy on small or occluded targets, while two-stage designs spend additional computation on refinement.[S01]

The limitation becomes sharper for long, narrow fabric faults. In the optimisation study, YOLOv4 performed poorly on long and narrow defects, motivating separate treatment of classification and segmentation. This is valuable negative evidence. It shows why a general detector trained around common natural-image shapes should not be accepted solely because its aggregate benchmark is respectable. The local test set must contain enough elongated warp and weft faults to expose this failure.[S02]

Pruning removes redundant structure and retrains the remaining network. The UNet++ optimisation study uses Bayesian optimisation to tune pruning parameters and reports both inference reduction and post-processing accuracy. Its results are useful here because they retain the less convenient outcome: image-level accuracy reaches 92.75, 94.87 and 95.60 percent on two self-made fabric datasets and a knitting dataset, but 81.82 percent on the nanofibre dataset. Selective reporting of only the first three would overstate generality.[S02]

Exhibit 23 keeps those four outcomes on one scale. The comparison concerns image-level accuracy after pruning and post-processing, not detector mAP, and the datasets are not equally difficult.

The exhibit supports two conclusions. Compression can preserve high task accuracy on fabric data, which matters when a detailed model must run near the line. It can also expose a domain boundary rather than remove it. The correct procurement response is to request the equivalent per-dataset results for the factory's own fabric families, together with precision and recall. A single average across unlike sets would hide exactly the variation that deployment needs to reveal.[S02]

Compression also changes the acceptance test. The unpruned and pruned variants should be compared on the same images, hardware and preprocessing path. Any loss of critical-defect recall must be visible before latency savings are considered. The later edge-engineering chapter examines device throughput in more depth. At the architecture stage, model size is only a hypothesis about deployability, while retained per-class performance is the evidence.