How does AI detect defects in fabrics and garments?
Computer vision is moving textile inspection from the human eye to cameras and models. How it works, where it is used, and why data and defect definitions matter most.
KEY TAKEAWAYS Summary by the editors
- AI quality control in textiles uses cameras and computer vision models to detect defects such as holes, stains, broken yarns and stitching faults on fabric rolls or finished garments.
- Deep learning methods, especially convolutional neural networks, now make up the majority of research approaches to automated fabric defect detection, according to a 2024 survey in Electronics.
- The main obstacles are a lack of shared defect definitions, mostly private training datasets and the computing power needed on factory floors.
- Automated inspection works best on uniform fabrics at the mill stage; inspection of complex garments still relies heavily on trained human inspectors.
- Better quality data matters more under EU rules that prioritise durable products and restrict the destruction of unsold goods.
AI detects defects in fabrics and garments by analysing camera images with computer vision models trained on examples of good and faulty material. The system flags holes, stains, slubs, broken yarns or stitching faults so that inspectors can check them, and it records each defect as data. It is most mature for fabric inspection at the mill; for finished garments it usually supports rather than replaces human inspectors.
Why does quality inspection matter in fashion?
Defects found late are expensive. A flaw that is missed on the fabric roll can end up in dozens of cut pieces, then in finished garments, then in customer returns. Traditional inspection relies on trained people examining fabric on inspection frames or checking garments by hand, which is demanding work where attention naturally drops over a shift.
European policy is also raising the stakes for product quality. The EU Strategy for Sustainable and Circular Textiles sets the vision that by 2030 textile products on the EU market should be durable, repairable and recyclable. Better inspection data helps brands understand where quality problems originate and address them with suppliers.
How does AI fabric inspection work?
- Image capture: line-scan or area cameras with controlled lighting photograph the fabric as it moves over an inspection machine, or garments at a station.
- Detection: a model analyses each image and marks regions that deviate from normal texture or pattern.
- Classification: the system assigns a defect type and severity, often based on the company's grading standard.
- Human review: an inspector confirms or rejects the finding; confirmed and rejected cases can be used to retrain the model.
- Data output: each defect is logged with position and type, creating a quality map of the roll or batch.
A 2024 survey of computer vision approaches published in the journal Electronics found that deep learning and machine learning approaches now make up the majority of methods in the field. Convolutional neural networks dominate, including one-stage detectors from the YOLO family and two-stage detectors such as Faster R-CNN, while generative models such as autoencoders are used to learn what normal fabric looks like and flag anything unusual. Earlier research, summarised in a widely cited 2008 survey in IEEE Transactions on Industrial Electronics, relied on statistical, spectral and model-based methods, and noted that defects are difficult to detect because they are vague and fall into many classes.
Which defects can AI detect?
| Task | Examples | Suitability for AI today |
|---|---|---|
| Plain and uniform fabrics | Holes, stains, broken ends, slubs, thick places | High, provided training images exist |
| Patterned and printed fabrics | Misprints, pattern faults, colour shading | Medium, the pattern itself can be mistaken for a defect |
| Shade and colour consistency | Shade variation between rolls or panels | Medium, depends on calibrated lighting and cameras |
| Finished garments | Seam puckering, skipped stitches, misaligned panels | Emerging, garments are three-dimensional and vary by style |
| Measurements and fit | Garment dimensions against specification | Emerging, needs controlled positioning |
What are the limits of AI quality control?
The survey in Electronics identifies several obstacles that remain relevant for brands and suppliers considering automated inspection.
- No shared defect taxonomy: the authors note that there is still no consensus on which defects should be considered, so models trained in one mill may not match another's standards.
- Private data: most research uses private datasets, and public benchmarks are small or outdated, which makes it hard to compare claimed accuracy.
- Computing resources: deep learning can require hardware that is impractical in factories with limited infrastructure.
- Rare defects: models learn from examples, and the rarest faults are by definition hard to collect.
- Changing products: every new fabric, colour or print can shift what normal looks like, requiring retraining or recalibration.
What does an AI inspection system need to work well?
The technical set-up is only part of the task. Stable, calibrated lighting and cameras matter, because changes in light can look like shade variation or texture faults to a model. Above all, the system needs labelled examples: images of each relevant defect type, graded according to the standard the company actually uses, plus many images of acceptable material so that normal variation is not flagged.
Agreeing on that standard is often harder than expected. Mills, garment factories and brands may grade the same flaw differently, and what counts as acceptable can vary by product category and price level. Writing down defect definitions and severity levels before training a model avoids building a system that is precise but measures the wrong thing. Integration is the final step: inspection results are most useful when they reach the systems where decisions are made, for example to hold a roll, adjust cutting or raise a claim with a supplier.
Who should own AI quality data in the supply chain?
Most inspection takes place at mills and garment factories, not at brands. That raises practical questions: who pays for the equipment, who owns the images and defect logs, and how results are shared. Brands that want to use inspection data for supplier evaluation or durability claims need agreements on data formats, access and retention with their suppliers.
Linking defect data to product and order data has a further benefit. Patterns that show up in returns, such as seams failing on one style, can be traced back to batches and suppliers. This kind of connection is also useful as companies prepare product information for the future Digital Product Passport.
How should a fashion company get started?
A sensible starting point is a single, high-volume quality problem, such as fabric defects found at cutting, with a clear baseline of how many are found and how many slip through. A pilot should run alongside existing manual inspection, so that results can be compared on the same material. Human inspectors remain essential: they define the standards, confirm findings and handle the cases the model has not seen before. The aim is fewer missed defects and better data, not inspection without people.
Frequently asked questions
Can AI replace human quality inspectors in textiles?
Not completely. AI can scan fabric consistently and flag likely defects, but people are still needed to define standards, confirm findings and handle unusual cases, especially for finished garments.
How accurate is AI fabric defect detection?
Accuracy depends heavily on the fabric type, the defects and the training data. Research results are mostly based on private or small datasets, so companies should test systems on their own materials before relying on published figures.
What technology is used for automated fabric inspection?
Systems use industrial cameras with controlled lighting and computer vision models, today mostly convolutional neural networks. Some approaches learn what normal fabric looks like and flag anything that deviates.
Is AI inspection used for finished garments?
It is emerging. Garments are three-dimensional and vary by style, which makes automated inspection harder than for flat fabric, so it usually supports human inspectors rather than replacing them.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
