An AI fabric inspection setup that works well on one fabric can miss defects or flag good fabric as soon as production moves to another. A hole in a plain woven fabric, a hole in a jersey knit and a hole in a nonwoven web are the same defect on paper, but each sits on a different background, under different light and with different fabric movement. For the quality manager, the production manager and the automation engineer who keeps the line running, the question is the same: can the same camera, light and AI model handle the new fabric? Sometimes.
The camera may stay as it is, but the lighting, the fabric handling and the model can all behave differently on a new fabric, so each one needs a test on representative samples. What decides the answer is how the new fabric differs from the current one in structure; in the way it handles light and in the way it moves.
Those differences show up in five places: imaging, lighting, fabric handling, the AI model and the acceptance criteria. Working through them in that order shows which parts of an existing setup can carry over to a new fabric, which need new settings, and what has to be tested before the fabric is approved.
What Changes When You Switch Fabric Types?
The structure of the fabric changes first, and the structure decides what the camera sees as normal.
Woven: interlaced yarns form a regular grid, so faults show as breaks in a repeating pattern.
Knitted: interlocking loops let the fabric stretch, and weft knits tend to curl at the edges [3], so the structure in the image can shift with handling.
Nonwoven: a bonded fiber web, defined in its own standard vocabulary [11], has no repeating structure, so the normal image is random and uneven.
The defect list changes too. Holes and oil stains can appear in both woven and knitted fabric, but broken ends, broken picks and missed picks belong to weaving, while needle marks, set-up faults and missing yarns are common in knits [3].
Standard terminology for fabric defects covers both woven and knit fabrics, which is a useful shared reference when you list the defects for each fabric [9].
Within each fabric type, five properties often decide how the defect shows in the image:
Transparency: thin or open fabric lets light pass through, so the background behind the fabric becomes part of the image.
Hairiness and lint: loose fibers on the surface add fine texture and can look like small marks.
Stretch: elastic fabric can change shape as it moves, which changes the size and spacing of the structure in the image.
Pattern: a printed or woven design adds planned variation that the system must tell apart from defects.
Luster: shiny or coated surfaces reflect light, which can hide a defect or create bright spots that look like one.
Each property can push the same setup in a different direction. That is why a new fabric needs its own check, even when the line, camera and software stay the same.
Can the Same Defect Look Different on Different Fabrics?
Yes. The defect may be the same, but the background, contrast and light path change. Figure 1 shows a hole, a stain and a structural fault on three structures. In the woven grid, the hole breaks a regular pattern. In the knit, the loops around it can pull open. In the nonwoven, the hole sits in a random web next to areas that are thinner by nature.
The way the fabric is lit changes the picture as well. With light from behind, a hole passes more light and shows as a bright spot, while a thick area passes less light and shows darker [7].
In a nonwoven lit from behind, areas with more fiber look darker and areas with less fiber look lighter [5], so the system has to separate a real hole from the normal variation in the web.
A stain and a structural fault behave differently again. A stain changes the color or brightness of the surface, so it stands out on a plain, light fabric but competes with the background on a dark, patterned or naturally uneven one. A structural fault, such as a missing pick in a woven or a needle line in a knit, breaks the repeat of the structure and stands out against a regular background. A nonwoven has no repeat to break, so a thick spot has to be judged against the normal density range of the web.
Patterns raise the difficulty again. Research on patterned fabrics describes the variety of pattern and texture as a major difficulty for defect detection [6]. A stripe, check or jacquard motif is planned variation, and the system needs enough good samples to learn what is normal before it can judge what is not.
Good fabric can also carry marks that are not defects. A dataset recorded on circular knitting machines labels normal production marks, called cutlines, separately from real defects, so that models can learn to keep them apart [2].
Can You Keep the Same Camera and Lens?
Often you can, but check four points on the new fabric before you decide. The aim is to confirm the image, not to choose new hardware.
Defect resolution: is the smallest defect that matters on the new fabric still large enough in the image to be seen and measured?
Focus: a thicker, raised or lofty fabric can sit at a different height, which can move the surface out of focus.
Distance and angle: a change in fabric path or thickness can change the working distance between lens and fabric.
Visible area: check that the full width, including the edges, stays in view. Curling edges on knits and changes in width can push part of the fabric out of the image.
Fixed settings matter here. In the circular knitting dataset above, the camera used an 8 mm lens with a manual focus ring, and the corners of the raw images showed slight shadowing caused by the camera mount and the lighting around the knitting zone [2]. Settings like these are chosen for one fabric and one position, so they need checking when either changes.
If the defect is too small, out of focus or partly outside the view, the camera position or optics need a fresh test before the AI model is judged. The camera types used for computer vision in fabric inspection, from area-scan to line-scan, are compared in a separate guide.
When Does a New Fabric Need Different Lighting?
Lighting is usually the first part of the setup to recheck, because fabrics differ in how they reflect and transmit light. Light from the front shows how the surface reflects, while light from behind shows how much light passes through. One research team captured both views of 38 fabrics because they give complementary information [1].
The fabric decides which view works better. An illumination patent for web inspection uses light from behind for material that is not opaque, and light reflected from the front for opaque material [8].
Each kind of change points to a different fix:
Color: a darker fabric reflects less light, so intensity or exposure may need to rise. Color also changes contrast, so a defect that was clear on white fabric can be faint on black. One patented system for knitted fabric uses a lit panel behind the web as a constant contrast reference, because finished fabric can be any color from bleached white to black [7].
Transparency: once light passes through the fabric, the background and anything behind it enter the image. Moving from front to back lighting, or adding a new background, is a new arrangement and needs a full test.
Hairiness: loose fibers scatter light and can show as small marks, especially under light at a low angle. Check whether a different angle reduces them, or whether the model has to learn them as normal texture.
Reflection: a shiny or coated surface can create glare that hides a defect or looks like one. Diffuse light or a new angle changes the arrangement, not just the intensity.
Sometimes an adjustment is enough. In one patented fabric inspection system, the allowed range of light intensity is adjusted to the type of fabric being monitored [7].
In nonwoven imaging with light from behind, exposure time has to be chosen carefully to avoid areas that are overexposed [5].
When the light changes and the model does not, results can drop sharply. In the circular knitting dataset, a model scored an average F1 of 86.84% when the test images had lighting similar to the training images, and 48.93% when the lighting differed [2]. F1 combines missed defects and false alarms into one score, and higher is better. The figures come from one dataset, but they show why lighting has to be tested on the new fabric.
Use this rule of thumb. If the new fabric only needs more or less light, adjust and retest. If it needs light from a different direction, a different angle, or light from behind instead of the front, test the whole lighting arrangement again. Lighting is the first step of the imaging chain. The steps that follow, from preprocessing to the final decision, are explained in fabric defect detection using image processing.
Do Fabric Tension and Handling Settings Need to Change?
They often do, especially when moving between woven and knitted fabric. The camera sees the fabric as it is handled, so four handling effects show up in the image:
Tension: too much tension stretches the structure and changes the spacing of yarns or loops, and uneven tension distorts it locally.
Wrinkles and creases: a fold casts a shadow and forms a line across the fabric, which can look like a linear defect such as a missing pick.
Movement: if the fabric lifts, flutters or drifts sideways, its distance from the lens and its position in the view change, which can blur the surface or push an edge out of the image.
Deformation: knitted fabric can curl at the edges and lose its shape under handling, so the structure in the image no longer matches the approved reference.
Distortion can look like a defect. In the circular knitting dataset, fabric deformed by tension imbalance, pulling or compression is recorded as its own defect class, because twisting and local thickness changes alter the surface [2].
Handling can also cause false stops. In a patented system for inspecting tubular knitted fabric, the fabric tended to narrow in the middle, an hourglass shape, which could cause a false stop because the cameras might read it as a defect passing through the view [7].
When a new fabric arrives, run it at its normal tension and speed, then check the images for stretch, wrinkles, edge curl and movement across the width. Changes to the fabric path or tension should be confirmed with production, because they also affect the fabric itself.
What to Recheck for Each Fabric Type
The table brings the imaging and handling points together. Use it as a checklist for samples of your own fabric, not as a final answer.
| Fabric type | Imaging issue | Setting to check | Test needed |
|---|---|---|---|
| Plain woven | Regular grid; small faults such as broken ends and missing picks | Resolution, focus, light angle | Images of each target defect at the smallest size that matters |
| Patterned woven (stripes, checks, jacquard) | The planned design can look like a defect | Light that shows both pattern and defect; normal samples of the pattern for the model | Good samples across the full pattern repeat and each colorway |
| Knitted | Stretch, edge curl and loop structure; knit-specific faults | Fabric tension and path, edges in view, focus | Images at running tension, including edges and distorted areas |
| Nonwoven | Random structure; normal density variation | Light from behind, light intensity, exposure | Normal thin and thick areas compared with real defects |
| Sheer or open fabric | Light passes through; the background shows | Front or back lighting, background behind the fabric | The same samples under both lighting options |
| Shiny or coated fabric | Reflections can hide defects or look like them | Light angle and diffusion | Samples imaged at several positions across the width |
| Dark or strongly colored fabric | Less light reflected; lower contrast for some defects | Light intensity and exposure; a lit panel behind the fabric as a contrast reference | Target defects imaged on the darkest and lightest colors you run |
| Brushed or hairy fabric | Loose surface fibers can look like small marks | Light angle | Good samples checked for false alarms caused by surface fibers |
Swipe to see the full table
Can One AI Model Inspect Multiple Fabric Types?
It can, but only if testing shows that it performs well enough on each fabric. There are three common options:
One shared model: the same model inspects several fabrics. This is simplest to run, but every fabric must be covered by the training data and the test.
A shared model with fabric-specific settings: the model stays the same, while settings such as decision thresholds are adjusted for each fabric.
More data or a dedicated model: when the fabric differs too much, the model needs retraining with more images of it, or a model of its own.
Research shows why the choice needs data. In a study of 38 woven, patterned fabrics, a model reached a score of 95.2 on fabrics it had seen in training and 85.6 on a fabric held out of training, with wide variation between fabrics (plus or minus 14.7) [1]. The score is AUPR, and higher is better. The authors link the drop to the gap between the training fabrics and the new one. In the same study, retraining with defect-free images of the held-out fabric raised the median score to 91.8.
Methods built to work on new surfaces still need some normal images of that surface. A study of anomaly detection across different textures used a small number of normal images from the new texture, and the authors expect performance to drop when the new texture is very different from the training textures [4].
The practical question is whether the model handles your next fabric well enough, and only results on that fabric’s samples can answer it.
How Do You Define Acceptable Variation for Each Fabric?
Before testing, write down what “good” means for the new fabric. Without that, a false alarm and a missed defect cannot be counted. Define three things:
Reference sample: approved fabric that shows the normal look, including normal variation in pattern, texture and density.
Target defects: the defects that matter for this fabric, with the smallest size that matters for each.
Acceptance limits: what is accepted, what is rejected and what goes to an operator for review.
Acceptance usually depends on the customer as well as the plant. Standard methods for visually inspecting and grading fabrics can be used to accept rolls and shipments when the buyer and seller agree [10].
Point systems show how limits are applied in practice. Under the four-point system, a published study on knitted fabric assigns penalty points by defect length, gives holes the maximum points, and compares points per 100 square yards with the acceptance level [3]. Whatever system you use, the inspection setup has to reflect the same limits.
What Should You Test Before Approving a New Fabric?
Test the new fabric on the line, with samples that represent normal production. A useful test set includes:
Good samples across the normal range of the fabric, including pattern repeat, color and batch.
Defective samples of each target defect, including the smallest size that matters.
Borderline samples that sit near the acceptance limit, where people may also disagree.
Measure two results separately. Missed defects are defective areas the system passes, and false alarms are good areas it flags. A single overall score can hide one of them. Count missed defects against the defective samples, and false alarms against the good samples or the length of fabric inspected, so both can be compared with the acceptance limits.
Run the test under real production conditions: the normal line speed and tension, the actual light at the inspection point, the full fabric width, and more than one run or batch. In the circular knitting dataset, models scored an average F1 of 39.58% on fabric produced at lower speeds than the fabric in their training data [2]. Conditions that differ from the setup used in training can change the result, even on the same fabric.
Until the new fabric is approved, keep the current inspection running in parallel. This lets you compare the results of the new setup against it, and no fabric is released on an unproven setup.
Approve the fabric only when both results meet the limits you defined. If they do not, the result points to the part of the setup to fix next: the camera, the light, the handling or the model.
What Should Be Saved for Each Fabric's Inspection Setup?
Save the approved setup for each fabric, so the same test can be repeated and the same conditions restored at each changeover, when the fabric returns to the line. A useful record includes:
Imaging settings: camera position, lens, focus, exposure, and the type, position and intensity of the light.
Handling settings: fabric path and tension during the test.
Model version: the model and the decision settings that were approved.
Target defects and acceptance limits: as defined for this fabric.
Test results: missed defects and false alarms, with the date and the samples used.
ISO 9001:2026 includes requirements for documented information and for the control of changes in production [12], so the record can sit inside the quality system you already run. Keep it there whether or not the inspection software also stores its own settings.
At each changeover, check the restored setup before production starts. Run a few reference and defective samples through it and confirm that the results match the approved test.
When Is a Shared Inspection Setup No Longer Enough?
A shared setup stops being enough when the test results show that one fabric needs conditions that would harm another. Watch for these signs:
A target defect is not visible on the new fabric under the shared light, camera or focus.
The new fabric needs light from behind while the existing fabrics need light from the front, or the reverse.
Settings that reduce false alarms on one fabric increase missed defects on another.
Handling the new fabric creates distortion that the shared path cannot avoid.
The model stays below the agreed limits on the new fabric after more data has been added.
At that point, check dedicated imaging settings for that fabric or a separate inspection station. The decision rests on the test results for each fabric. A line that runs many fabrics can still share one setup if every fabric passes, and a line that runs two can need separate settings if one of them fails.
AI-Innovate assesses this question on your own fabrics and target defects, not from a product list. AIxEye, its 2D automated visual inspection system, is used for woven and knitted fabric inspection. AIxLight, its engineered machine vision lighting, is used where low-contrast defects, reflective materials or variable textures limit what the camera can see. Whether either fits a given fabric is decided by imaging tests on that fabric. The textile inspection applications AI-Innovate supports show where these fit across the textile process.
Share representative samples of each fabric you run, both acceptable and defective, along with the target defects for each. AI-Innovate will review whether one shared inspection setup is feasible across them, or which fabrics may need their own settings or station. Compatibility with every fabric or a specific result is not promised before testing.
References & Sources
- Rippel, O., Zwinge, C., Merhof, D. Increasing the Generalization of Supervised Fabric Anomaly Detection Methods to Unseen Fabrics. Sensors, 22(13), 4750, 2022. doi:10.3390/s22134750
- Ni, Y.-Q., Huang, P.-K., Wang, W.-J., Liang, D., Lin, C.-Y. TSFabrics: A Time-Series Fabric Dataset for Real-Time Defect Detection on Circular Knitting Machines. Scientific Data, 13, 379, 2026. doi:10.1038/s41597-026-06748-9
- Das Gupta, A., Sadek, Z., Hossain, M.S., Toha, T.R., Mondol, A., Habiba, S.U., Alam, S.M.M. An approach to automatic fault detection in four-point system for knitted fabric with our benchmark dataset Isl-Knit. Heliyon, 10(17), e35931, 2024. doi:10.1016/j.heliyon.2024.e35931
- Chen, S.-F., Liu, Y.-M., Lin, C.-C., Chen, T.P.-C., Wang, Y.-C.F. Domain-Generalized Textured Surface Anomaly Detection. IEEE International Conference on Multimedia and Expo (ICME), 2022. doi:10.1109/ICME52920.2022.9859637
- Moll, P., Wang, S., Coutandin, S., Fleischer, J. Analysis of Basis Weight Uniformity Indexes for the Evaluation of Fiber Injection Molded Nonwoven Preforms. Autex Research Journal, 21(3), 2021, pp. 341 to 351. doi:10.2478/aut-2020-0039
- Di, L., Long, H., Liang, J. Fabric Defect Detection Based on Illumination Correction and Visual Salient Features. Sensors, 20(18), 5147, 2020. doi:10.3390/s20185147
- US Patent 6,259,526 B1. Operator-free fabric web inspection system. Issued July 10, 2001. Patent text
- US Patent 7,382,457 B2. Illumination system for material inspection. Issued June 3, 2008. Patent text
- ASTM International. ASTM D3990-12(2020), Standard Terminology Relating to Fabric Defects. Standard page
- ASTM International. ASTM D5430-26, Standard Test Methods for Visually Inspecting and Grading Fabrics. Standard page
- ISO 9092:2026, Nonwovens: Vocabulary. Standard page
- ISO 9001:2026, Quality management systems: Requirements. Standard page