Quality control is one of fashion’s most persistent operational problems — and one of its least glamorous. Every garment that leaves a factory carries the invisible risk of a defect that will not be discovered until it reaches a customer, triggers a return, and generates a complaint. The cost of that journey — the return shipping, the processing, the lost customer, the brand damage — dwarfs the cost of catching the defect at the source.

For most of fashion history, quality inspection was entirely manual: trained inspectors examining garments under standardised lighting, checking seams, checking colour consistency, checking construction at a rate of perhaps one every few minutes. At scale, this is both slow and inaccurate. Human inspection accuracy for fabric defects runs at 70–80% under ideal conditions — and degrades significantly with fatigue, poor lighting, and the repetitive nature of the work.

Computer vision — the application of deep learning to visual inspection tasks — is transforming this. Systems that can examine fabric at resolutions the human eye cannot achieve, at speeds no human team can match, and with accuracy that does not degrade over a 12-hour shift, are now commercially deployed across fashion’s global supply chain. This article explains exactly how they work, where they are being used, what they are catching, and what remains beyond their current reach.

A computer vision quality control system is, at its core, a convolutional neural network (CNN) — a class of deep learning model specifically designed to process visual data — trained on a large labelled dataset of defective and non-defective samples. The model learns to recognise the visual signatures of specific defect types: the irregular weave pattern of a broken warp thread, the colour band of a dye lot inconsistency, the puckering shadow of a misaligned seam.

In a production deployment, fabric passes continuously through a camera system — often multiple high-resolution cameras at different angles and under controlled illumination — while the model processes each frame in real time. When the model detects a pattern that matches its defect signatures above a confidence threshold, it flags the location on the fabric, annotates the defect type, and triggers an alert or automatic stop mechanism on the production line.

The most advanced systems operate at multiple stages of production simultaneously — inspecting raw fabric at the weaving or knitting stage, checking cut panels before sewing, and performing final garment inspection before packing. Each stage catches different defect types, and catching defects earlier is dramatically cheaper: a flaw identified at the fabric stage costs a fraction of the same flaw identified in a finished garment.

The financial case for computer vision in quality control is built on one foundational principle: the cost of a defect compounds with every stage it passes through undetected. A thread flaw caught at the loom costs a few pence to fix — trim the roll, restart the weave. The same flaw caught after dyeing costs the dye batch. Caught after cutting, it wastes the cut panels. Caught in a finished garment, it wastes the entire labour and material cost of construction. Caught by the customer, it costs the product, the return processing, and potentially the customer relationship.

Industry data from Uster Technologies, which supplies quality measurement systems to the global textile industry, suggests that the cost of a fabric defect detected at the yarn stage is approximately 1x. The same defect detected at the finished garment stage costs 100–500x more to rectify — and a defect that reaches the end customer costs an order of magnitude more again once brand damage and returns processing are included. This 1:100:1000 cost ratio is the economic foundation on which the entire computer vision QC market is built.

The computer vision QC market in fashion has matured rapidly. A handful of specialist vendors dominate, each with slightly different approaches to the core inspection problem:

Inspectorio — one of the leading platforms in fashion supply chain QC — uses a combination of in-factory camera systems and mobile inspection tools to run AI-powered quality checks across its client network. The system builds a quality data layer across thousands of audits, enabling predictive models that can identify which factories, product types, and production runs carry the highest defect risk before inspection even begins. Clients include major fashion retailers and brands sourcing from Southeast Asia and South Asia.

QualityLine uses IoT sensors and computer vision at sewing machines to capture real-time production data, detecting stitch density deviations and construction anomalies as they occur rather than after the garment is complete. Their system allows factory floor supervisors to see defect rates by machine, by operator, and by time of day — identifying patterns that reveal the root cause of quality issues rather than just catching their outputs.

Uster Technologies has been the global standard in textile quality measurement for decades. Their USTER QUANTUM and USTER TESTER systems use AI-enhanced sensors to measure yarn evenness, imperfections, hairiness, and foreign fibre content at spinning mill speeds — catching quality issues at the earliest possible stage. Their AI enhancement, introduced over the past five years, has added predictive classification capabilities that go beyond measurement to recommendation: flagging specific spinning parameters to adjust before a quality drift becomes a defect.

On the brand side, PVH Corp (Calvin Klein, Tommy Hilfiger) has deployed computer vision inspection across its key Tier 1 supplier factories, integrating defect data directly into its supplier scorecards. Inditex runs AI-assisted final inspection at its logistics hubs, catching construction defects in finished garments before they enter the retail network. Levi Strauss uses computer vision to monitor the laser finishing process on denim — ensuring precise, repeatable distressing patterns without the chemical waste of traditional manual finishing.

The first generation of computer vision QC was reactive: cameras watched production and flagged defects as they appeared. The current generation is moving toward something more powerful — predictive quality: using the patterns in inspection data to anticipate where defects are likely to occur before they do.

Predictive quality models ingest time-series data from sensors embedded in production equipment — loom tension readings, needle pressure, motor temperature, humidity levels in the dyeing bath — alongside the inspection outcomes from finished goods. Over time, they learn the correlations between production conditions and defect rates: this loom produces warp breaks when humidity drops below 45%; this dye machine produces shade variation when the bath temperature varies by more than 2°C; this sewing machine produces skip stitches in the last two hours of a shift when its needle has not been changed.

Armed with these correlations, the system can intervene preventively — alerting operators to adjust machine parameters, schedule needle changes, or check humidity control before the defect rate rises — rather than waiting to catch defects after they have been created. Toyota’s manufacturing philosophy calls this jidoka: the principle that quality should be built in, not inspected in. Computer vision AI is finally making jidoka achievable at scale in fashion manufacturing.

Fabric defects are only one dimension of fashion quality. A garment can be flawlessly constructed and still fail quality control if its dimensions deviate from the specification — a sleeve 5mm too short, a waist 2cm too wide, a collar stand at the wrong angle. Dimensional accuracy is the second major domain of computer vision quality control, and it is in some ways more technically challenging than defect detection.

Traditional dimensional inspection requires a human inspector to lay a garment flat, position it against a measuring template, and manually check key points of measure against the technical specification — a process that takes several minutes per garment and is subject to the same human variability as any manual inspection. AI-powered 3D scanning and photogrammetry systems can now complete the same measurement in seconds, with millimetre-level accuracy, without requiring the garment to be laid flat.

Tukatech and Gerber Technology (now Lectra) have both integrated AI measurement verification into their cutting room systems — checking cut panel dimensions against digital patterns before sewing begins. Browzwear and CLO3D, primarily known as 3D design tools, are extending their technology into fit verification systems that can check a finished garment’s actual measurements against the original digital spec. The gap between the design specification and the physical outcome — which can compound through every step of the manufacturing process — is becoming measurable, trackable, and correctable in real time.

Computer vision does not stop at the factory gate. A fast-growing application is using the same technology to process the returns that escape factory inspection — assessing condition, catching missed defects, and making intelligent routing decisions about what happens to each piece.

Fashion’s return rate runs at 20–40% for online purchases. Each returned item must be assessed: is it resaleable at full price? Does it need repair? Is it suitable for secondary market? Does it have a manufacturing defect that should trigger a supplier quality claim? Made manually, these decisions take several minutes per item. Computer vision systems make them in seconds.

ASOS has piloted automated returns grading at its Barnsley fulfilment centre, using computer vision to assess condition across four criteria — original tags, visible damage, staining, and fabric condition — and automatically route each item. Optoro’s REDO platform reduces average time from return receipt to relisting from 14 days to under 48 hours. Every day a returned item sits in a processing queue is a day it is not generating revenue.

Any honest discussion of computer vision QC must address labour displacement directly. Fashion manufacturing in the Global South employs hundreds of millions of workers, a significant proportion performing manual inspection tasks. Computer vision automation raises legitimate questions about the impact on these workers.

The industry’s standard response — that automation creates new higher-skilled roles — is partially but not entirely accurate. Computer vision systems do require skilled operators, data annotators, and maintenance engineers. These roles are typically better paid and less physically demanding. But they require different skills, and the transition is neither automatic nor frictionless for workers whose experience is solely in manual inspection.

The honest assessment: computer vision QC will reduce manual inspection roles over time, though more slowly in labour markets where the human–machine cost differential is smaller. The responsible path for brands is to include supplier transition support, retraining investment, and workforce transition planning in deployment contracts — not as afterthoughts. Several major brands are beginning to add automation transition requirements to their supplier codes of conduct, a meaningful if nascent step.

For all its capabilities, computer vision quality control has genuine limitations that the industry should understand clearly — both to avoid over-reliance on automation and to direct R&D investment toward the remaining hard problems.

Tactile quality remains largely out of reach. The hand-feel of a fabric — its softness, drape, weight, and texture — is a primary component of perceived quality in fashion, and one that no camera system can currently assess. Sensors that measure specific physical properties (tensile strength, pilling resistance, friction coefficient) exist, but replicating the holistic tactile judgement of an experienced quality inspector has not been achieved and is likely many years away.

Novel defect types challenge models trained on known defect categories. Computer vision systems learn from labelled historical data. A new defect type — introduced by a new fabric, a new dye process, or a new construction technique — will not be detected by a model that has never seen it. Human inspectors, by contrast, apply general visual intelligence and contextual judgement that allows them to flag anomalies they have not specifically been trained to recognise. Keeping model training data current with evolving production processes is an ongoing operational requirement.

Ethical and aesthetic judgements — whether a particular variation in hand-printing makes a garment characterful or flawed, whether a slight colour deviation in a vintage-washed fabric is a defect or a feature — remain human calls. The boundary between quality control and aesthetic interpretation is blurry in fashion in a way it is not in, say, semiconductor manufacturing. Computer vision draws that line based on its training data; the nuanced calls still require human review.

Five Actions for Fashion Quality Leaders