5 fabric defect categories costing textile manufacturers the most—and how AI vision catches them at speed
270 items per hour. That is the throughput one HyperQ AI Vision customer — an automotive parts manufacturer running 8,000+ product variants — reached after replacing a hardware-locked vision platform. Manual visual inspection on the same part: 40 items per hour. The previous hardware-bundled microscope system: 60 items per hour. A 6.75x improvement, achieved not by a faster human and not by a faster microscope, but by inspecting every part — instead of sampling.
That distinction — sampling versus continuous coverage — is the structural gap inside every textile quality budget. For mills in Malaysia and Singapore supplying international garment and technical-textile buyers, it is the chargeback risk sitting inside every shipment that passed AQL.
This post does three things. First, it walks through the five fabric defect categories that escape sampling-based inspection most often: weaving, knitting, finishing, dimensional, and surface contamination. Second, it explains why model-based AI vision on line-scan cameras is the only inspection architecture that closes the coverage gap rather than just speeding up the sampling. Third, it grounds the architecture in three real production deployments — none of them textile, all of them directly applicable to fabric inspection.
What manual fabric inspection actually covers
A typical mid-sized woven line runs at 60-100 metres per minute and produces around 20,000 metres a day. One inspector at the end of that line is doing continuous visual processing of a moving web. The task is bounded by physiology, not motivation.
Inspector performance at hour 7 of an 8-hour shift on a moving fabric web is not the same as at hour 1. Training does not solve it. Supervision does not solve it. The cognitive load of repetitive visual processing produces measurable attention drop, and the drop accelerates as the shift progresses.
AQL was designed around this constraint, not against it. It is a batch-acceptance framework that sets a statistical threshold for accepting a lot based on how many defects appear in a sample. For roll-level inspection on rolls running 100 metres or longer, an inspector examines a defined section, and the roll is passed if that section meets the defect limit.
At standard inspection levels on lots greater than 10,000 units, AQL achieves roughly 1.6% physical coverage of the output. AQL 4.0 for Major Defects can pass lots with up to a 7% defect rate inside the unsampled portion.
The 98.4% that is not physically examined is accepted by statistical inference. Defects between sampling points do not appear in the inspection report. They appear in the buyer's incoming check.
An inspection system can only detect what it physically examines. The detection rate is downstream of the coverage rate, and AQL coverage is structurally bounded.
The five defect categories that escape sampling
Each family behaves differently. Some are point events. Some run lengthwise. Some accumulate gradually across tens of metres. The common thread: standard sampling misses them at predictable, repeatable rates.
1. Weaving defects
Broken end, weft loop, float, stop mark, reed mark, double pick, missing pick.
Most run along the length of the fabric. A broken warp end can extend across 2-8 metres before the loom operator catches it. A float starts at the metre where a yarn lifts and continues until the weave structure recovers.
A defect that starts at metre 30 and ends at metre 35 either falls inside the inspector's sampled window or it does not. Position is luck. The defect is real either way; the inspection report only reflects whether the inspector happened to be examining that metre.
Line-scan inspection captures every metre of warp and weft as the fabric moves past the camera. A continuous defect is detected at its start, its length is measured, its position is logged against the encoder. The defect class is recorded — not the verdict on a sampled section.
2. Knitting defects
Dropped stitch, hole, ladder, broken needle line, course shift, tucked stitch.
Knitting defects often present as a vertical column running down the fabric. A ladder or needle-line failure can run for tens of metres before being caught at the loom. Knitted surfaces are also more visually variable than wovens, which is why rule-based inspection systems that handle wovens reasonably well tend to fail on knits.
Where surface variability is high, fixed-threshold detection breaks down. Rule-based logic handles predictable checks; model-based AI handles variability. Fabric is exactly the variability case.
A model trained on the specific knit construction learns the acceptable variation range for that fabric. Drift outside that range — a missing column, a ladder, a tucked stitch cluster — triggers an alert with position data the operator can act on.
3. Finishing defects
Shade variation, stop mark, calender mark, abrasion mark, yellow stain, colour staining.
Finishing defects emerge gradually across a roll. A shade band that develops slowly across 50 metres is below the detection threshold of any single sampling point. An inspector looking at metre 25 and an inspector looking at metre 75 may each judge their section in spec. But when the buyer compares metres 1 and 75 of the same roll under consistent reference lighting, the drift is unmistakable.
Fixed-threshold rule-based vision struggles for a related reason. Shade variation requires normalisation against a reference profile and the ability to flag gradual drift, not hard-edge contrast.
Model-based AI vision normalises against the production target colour, tracks shade across the entire roll, and flags drift as a continuous deviation rather than a binary pass-or-fail. The mill gets a record of where on the roll the drift began, which lets them correlate it with dye-bath data, fixation temperature, or finishing chemistry.
4. Dimensional defects
Width variation, weft skew, weft bow, length error, weight variance.
These are often outside the scope of visual inspection entirely. They get caught at lab QC after the roll is finished. Or worse, at the buyer's incoming inspection. By that point the rework cost is fully loaded with finishing, packaging, and shipping.
Line-scan systems with the right calibration measure width and skew per metre and detect deviations as they occur. The mill can intervene mid-run instead of discovering the drift after the roll is sealed.
5. Surface contamination
Oil spots, dirt, fly contamination, foreign fibre, machine oil staining.
Contamination defects are point events. Small, scattered, often near loom edges. Manual inspection at 60 m/min has a high miss rate on small contamination because the visual angular size at the inspector's distance is below their effective resolution at speed.
At line-scan resolutions of 0.1-0.5 mm per pixel, contamination at sub-millimetre scale is detected reliably. The mill gets per-metre contamination logs that often surface upstream maintenance issues — a leaking bearing, a worn carrier — long before they degrade visible quality.
Why rule-based vision fails on variable fabric surfaces
A common pushback: we already have automatic inspection. A vision system was installed five years ago.
The question is whether the system is rule-based or model-based. Fabric surfaces are continuously variable in ways that defeat rule-based detection.
Yarn count varies within tolerance across a warp. Dye batches shift colour slightly run to run. Surface tension varies with loom speed and humidity. A rule-based threshold set at commissioning for one yarn count and one colour is not the same threshold needed for the next dye batch, or for the same warp running at 5% higher tension.
Operators respond to rule-based false rejects the way operators always respond. They widen the tolerances. The practical effect: as the production run lengthens and the variables drift, the inspection system becomes less sensitive to real defects, at exactly the point in the run where defect risk from process drift is highest.
Model-based AI vision learns the acceptable variation range from production samples instead of depending on a fixed threshold. It flags deviations from what the fabric actually looks like, not from what it looked like at commissioning. That is the specific technical capability that distinguishes HyperQ AI Vision from rule-based inspection on textile applications.
Three deployments that prove the architecture
HyperQ AI Vision has not yet been deployed in a Malaysian or Singaporean textile mill. The 47 production contracts to date are concentrated in semiconductor valves and sockets, automotive parts, display panels, PCB inspection, plating, packaging, and laser-engraving verification. The defect classes across those deployments — irregular surface defects, dimensional drift, low-volume rare events, multi-material variability — map directly onto fabric inspection. Three of them are worth examining in detail.
Client A — 8,000 product variants, automotive parts
An automotive parts manufacturer running 8,000+ product variants on a multi-material plastic-and-metal part. The previous inspection setup: a hardware-locked vision platform with 2 cameras and 2 lights per station. Manual product line changes were required for every variant switch — labour-intensive, error-prone, and limiting the throughput ceiling.
HyperQ AI Vision integrated with the production equipment so that when the product changed on the line, the inspection program updated automatically. No manual switching. The hardware was reduced from 2 cameras and 2 lights to 1 camera and 1 light per station — a 50% hardware footprint reduction on a more capable system.
Measured outcomes: 99% detection rate, 11,520 units per day, throughput of 40 items per hour (manual visual) → 60 items per hour (the previous hardware-locked microscope) → 270 items per hour on HyperQ AI Vision. A 6.75x improvement over manual and a 4.5x improvement over the incumbent hardware-bundled platform. The customer expanded from the initial deployment to 6 lines.
Timeline from first meeting: demo video at 2 weeks. Contract signed at 1 month. Implementation completed 4 weeks after contract. On-site setup 2 days. Sales terms: 50% deposit, 50% on delivery, 1-year free maintenance.
The reason the customer switched: the incumbent vision platform could not integrate with production equipment for automatic recipe switching across 8,000 variants. HyperQ AI Vision could.
Client B — irregular small parts, 2D where 3D was proposed
A Korean factory of a Japanese precision parts manufacturer. The product was small and complex. Defect types were irregular. Multiple rule-based vision vendors had been unable to inspect it.
The competing proposals were 3D vision systems — higher cost, slower inspection, longer integration. HyperQ AI Vision selected the lighting and cameras to match the product geometry directly, refined the data for irregular defects, and achieved the inspection in 2D.
Timeline: proposal and quotation at 1 week after first meeting. Contract at 2 months. Implementation 2 months from contract. On-site setup 2 days. The customer also requested LOT-unit data management mid-project, which was delivered.
The architectural lesson for textile: irregular, non-uniform surface defects do not require 3D imaging. The right combination of illumination, line-scan resolution, and a model trained on the actual production surface is sufficient — at lower cost and higher throughput.
Client C — 1 defect per year, customer-driven retraining
A display panel manufacturer where defects occurred at a rate of 1-2 per year. The incumbent rule-based and large-dataset deep-learning vendors required 10,000+ training images per defect class. With 1-2 defects per year, that requirement was structurally unsatisfiable.
The HyperQ approach: generate initial defect data through a demonstration, deliver labelling and training tools to the customer, and let the customer continuously improve the model independently. The model was trained with minimal initial data and is now operated and upgraded by the customer on-site. Data export was disallowed and remote access restricted, so all training and tuning happens inside the customer's facility.
Timeline: contract at 2 months from first meeting. Implementation 1 month from equipment installation.
The architectural lesson for textile: a fabric defect that appears once every 50 rolls is functionally unreachable for a 10,000-image training requirement. HyperQ AI Vision needs roughly 1,000 images to train a working model — 10x less data, patented methods, customer-managed retraining as new defect classes emerge.
The hardware that enables 100% coverage
Line-scan cameras are the enabling technology. The relevant parameters for textile inspection:
- Resolution: 0.1-0.5 mm per pixel, sized to the smallest defect class the mill needs to catch.
- Line rate: synchronised to the encoder so every pixel of the moving web is captured.
- Illumination, chosen for the defect class — raking low-angle directional for surface texture and pulls; transmitted back-light for density variation, holes, and thin spots; coaxial on-axis for glossy surface defects and calender marks; fluorescent for contamination, oil, and foreign-fibre detection.
HyperQ AI Vision is hardware-agnostic. It works with any industrial camera brand — auto-detection on GigE — which means mills are not locked into a proprietary camera ecosystem and can specify hardware on cost and availability rather than vendor compatibility. The 30-50% hardware cost saving versus hardware-locked ecosystems is structural, not promotional.
What 100% coverage actually changes for a mill
A mill that inspects every metre has different data than a mill that samples. Three operational shifts show up in the deployment record across non-textile industries and translate directly to fabric.
Per-metre defect data. A defect is logged at the metre it occurred, not the section it was sampled in. Client A's system produces per-unit logs across 11,520 units per day; the textile equivalent is a per-metre log across 20,000 metres per day. Buyer disputes get resolved with position data instead of narrative.
Process traceability. A shade band at metres 47-52 can be correlated with the loom tension log, the dye fixation data, and the finishing chemistry at that timestamp. None of this exists with sampling-based inspection because the defect data does not exist for the unsampled portion of the roll. Client B's mid-project request for LOT-unit data management is the same mechanism applied to small precision parts.
Reduced chargeback exposure. A roll that passes AQL but ships with a defect cluster between sampling points produces a buyer-side rejection the mill cannot dispute on its own inspection records. With continuous coverage, the inspection report and the buyer's incoming check operate on the same data set.
Three evaluation questions before any fabric inspection deployment
Vendor answers vary widely on each, and the differences are operationally significant.
1. Does the system inspect every metre, or sample at intervals?
Line-scan cameras inspect every pixel as the fabric passes. Area cameras capture frames at fixed intervals, and the fabric between frames is not inspected. For roll-level coverage, those are not equivalent. Clarify the actual coverage model before evaluating any accuracy claims; 99% accuracy on 30% sampled coverage is materially different from 99% accuracy on 100% coverage.
2. How does the system handle shade variation across a dye run?
Fixed-threshold detection produces false positives on normal batch-to-batch colour variation and fails to flag gradual drift accumulating across 50+ metres. Ask specifically: does the system normalise against a reference profile, and does it flag gradual drift, or only hard-edge colour steps? The answer separates rule-based detection from model-based detection.
3. What does training data look like, and who owns the retrained model?
A vendor that requires 10,000+ training images per defect class is incompatible with mills running short batches and rare defects. A vendor that owns the retrained model and bills per retraining cycle locks the mill into the vendor's pricing. Ask: how many images to train a working model, who labels the data, and who controls retraining as defect classes evolve? HyperQ AI Vision uses ~1,000 images per defect class, supplies labelling tools to the customer, and runs retraining on the customer's premises.
The HyperQ AI Vision approach
Three differentiators carry across the 47 deployments and apply directly to textile inspection.
Low-data training. Roughly 1,000 production images to train a working model, against the 10,000+ requirement of rule-based and conventional deep-learning systems. Patented methods. For mills running short batches and frequent fabric changes, that is the difference between AI vision being feasible and not.
8,000+ product models, zero-configuration. Pre-built defect classes and product profiles from prior deployments are available as starting points. New product lines extend the system through prompt extension, not retraining cycles.
Universal camera compatibility. HyperQ AI Vision works with any industrial line-scan camera brand. No proprietary rigs. The 30-50% hardware cost saving versus hardware-locked ecosystems is structural — the software runs on the customer's choice of optics. Mills keep replacement and upgrade flexibility, and the system integrates with existing infrastructure rather than replacing it.
The principle running through all three: no vendor lock-in. Mills should not have to rebuild their production line around the inspection vendor.
What a Malaysia or Singapore mill deployment would look like
The deployment record across automotive, semiconductor, display, PCB, and packaging gives a reasonable baseline for a textile pilot. Demo and proposal in 1-2 weeks from a sample. Contract in 1-2 months from first meeting. Implementation in 4-8 weeks from contract. On-site setup 2 days. 50% deposit, 50% on delivery. 1-year free maintenance.
A textile pilot starts with a sample. Send a roll segment representative of the fabric class and the defect classes the mill cares about. The Hypernology team builds a demonstration model on that sample and shows the detection output before any contract conversation begins. If the demonstration does not meet the defect classes the mill specifies, the project does not move forward.
Full coverage is the structural difference between a quality programme and a sampling exercise. A mill that inspects every metre has different data, different traceability, and different chargeback exposure than a mill that samples — regardless of how skilled the inspectors are or how rigorous the AQL discipline is. The defects between sampling points ship under either inspection model. They only fail to ship under continuous coverage.
For textile manufacturers in Malaysia and Singapore evaluating their inspection model against international buyer requirements, the question is not whether AI vision is more accurate than a human inspector at any given moment. It is whether 100% surface coverage produces different quality outcomes than sampling. The 47 production deployments answer that already.
Send a sample
Send a roll segment representative of the fabric class and defect classes that matter on your line. HyperQ AI Vision builds a demonstration model on the sample and returns detection output within two weeks. No contract conversation until the demonstration meets your defect specification.