What is computer vision in fashion? From defect detection to shelf monitoring
Computer vision lets machines read images and video: spotting fabric faults, tagging product photos, powering visual search and checking shelves. What it does, what it needs and where its limits lie.
KEY TAKEAWAYS Summary by the editors
- Computer vision is a field of artificial intelligence that uses machine learning to process, analyse and interpret images and video, with tasks such as classification, object detection, segmentation and visual inspection.
- In fashion, computer vision is used along the value chain: fabric and garment inspection in production, automated attribute tagging of product images, visual search for shoppers and monitoring of store shelves.
- The WiseEye system developed at The Hong Kong Polytechnic University inspects fabric during weaving and was reported to detect about 40 common fabric defects at a resolution of up to 0.1 mm per pixel.
- Google said in October 2024 that Lens handles nearly 20 billion visual searches a month and that 20 percent of Lens searches are shopping-related.
- Vision systems depend on representative labelled images, controlled lighting and ongoing retraining, and camera use involving people raises data protection and EU AI Act questions.
Computer vision is the branch of AI that lets software interpret images and video. In fashion it already does concrete jobs: finding faults in fabric as it is woven, checking finished garments, tagging product photos with attributes such as colour and neckline, letting shoppers search with a picture, and checking whether store shelves and rails are correctly stocked. It works best on narrow, repeatable visual tasks with good images and well-labelled training data.
What is computer vision?
IBM defines computer vision as a subfield of artificial intelligence that gives machines the ability to process, analyse and interpret visual inputs such as images and videos, using machine learning to extract useful information. Rather than following hand-written rules, modern vision models learn patterns from many labelled example images, for instance thousands of photos of fabric marked "defect" or "no defect", and then apply them to new images.
What tasks can computer vision perform?
IBM lists a set of core tasks. The ones most relevant to fashion are:
- Image classification: assigning an image to a category, such as "dress" or "trousers".
- Object detection: locating items in an image with bounding boxes, for example each garment on a shelf.
- Image segmentation: labelling images pixel by pixel to find exact outlines, useful for cutting out products from backgrounds.
- Visual inspection: automatically detecting defects or faults.
- Optical character recognition: reading text such as care labels, price tags or packing lists.
- Object tracking and pose estimation: following items or body movement across video frames.
How is computer vision used along the fashion value chain?
| Stage | Application | Main prerequisite | Maturity |
|---|---|---|---|
| Textile production | Fabric defect detection during weaving or inspection | Fixed cameras, consistent lighting, labelled defect images | Established in pilots and production |
| Garment manufacturing | Checking seams, prints and measurements | Defined quality standards per style | Emerging |
| Warehousing | Reading labels, verifying cartons and returns | Standardised labels and packaging | Established |
| Product content | Auto-tagging attributes from product photos | Agreed attribute taxonomy | Established |
| E-commerce | Visual search and similar-item recommendations | Large, well-tagged image catalogue | Established |
| Stores | Shelf and rail monitoring, out-of-stock detection | Camera coverage, planograms, privacy controls | Pilots, mixed results |
How does AI detect fabric defects?
Fabric inspection has traditionally been done by people at inspection frames, a task that is tiring and error-prone over long shifts. Vision systems mount cameras over the fabric, use controlled lighting and compare each image against what normal fabric looks like. A documented example is WiseEye, developed by the textile and apparel AI research team at The Hong Kong Polytechnic University's Institute of Textiles and Clothing. According to PolyU, the system uses an LED light bar and a camera on a rail to capture the full fabric width during weaving, applies an AI-based machine vision algorithm to detect about 40 common fabric defects at a resolution of up to 0.1 mm per pixel, and provides real-time analytics and alerts.
The practical value of such systems lies in catching faults early, before fabric is cut and sewn. Their limits are equally practical: new fabrics, prints or textures look different from the training data, so models need new labelled examples and periodic retraining, and results depend heavily on stable lighting and camera calibration.
Further down the chain, the same techniques are applied to cut pieces and finished garments, for example checking print placement, stitching or colour consistency against a reference. This is harder than inspecting flat fabric, because garments are three-dimensional, soft and vary by size, so many factories still rely mainly on human quality controllers. In warehouses, vision is more mature: reading labels with optical character recognition, checking carton contents and grading returned items are well-defined tasks where a camera at a fixed station sees the same kind of object again and again.
How do visual search and image tagging work?
On the commercial side, vision models turn product images into numerical representations that can be compared. A shopper's photo of a jacket is matched against catalogue images to find similar items, and new product photos can be tagged automatically with attributes such as colour, pattern or sleeve length. Consumer behaviour shows why this matters: Google said in October 2024, as reported by Glossy, that Lens handles nearly 20 billion searches a month and that 20 percent of Lens searches are shopping-related. For brands and retailers, the lesson is that clean, consistent product imagery and attributes determine whether their products are found visually.
What about shelf monitoring in stores?
Shelf and rail monitoring uses fixed cameras, handheld devices or robots to compare what is on display with what should be there, flagging gaps, wrong placement or missing price labels. In fashion, the task is harder than in grocery because garments are folded, hung and handled by customers, and items differ by size and colour in ways cameras find difficult to distinguish. Combining camera data with RFID or inventory records can therefore be more reliable than relying on cameras alone. Business cases should be tested in pilots that measure staff time saved and sales recovered, not only detection accuracy.
What are the risks and requirements?
Three issues come up repeatedly. First, data: vision models need many representative, correctly labelled images, including rare defects or unusual products. Second, operations: cameras, lighting and integration with production or store systems cost money and need maintenance. Third, people and law: any system that captures workers or shoppers falls under data protection rules. The EU AI Act, in force since 1 August 2024, bans certain practices from 2 February 2025, including emotion recognition in workplaces and untargeted scraping of CCTV footage to build facial recognition databases. Fashion companies should design vision systems to look at products, not people, wherever possible.
Frequently asked questions
What is computer vision used for in fashion?
Computer vision is used for fabric and garment quality inspection, reading labels in warehouses, tagging product images with attributes, visual search in online shops and monitoring shelves in stores. It is most reliable on narrow, repeatable visual tasks with good image quality.
Can AI detect fabric defects better than humans?
Vision systems can inspect consistently without fatigue and at high resolution; the PolyU WiseEye system, for example, was reported to detect about 40 common defect types. Performance depends on lighting, calibration and training data, and new fabrics or prints usually require additional labelled examples.
How does visual search work in online fashion shops?
A vision model converts the shopper's photo and the catalogue images into numerical representations and returns the most similar products. Results depend on the quality and consistency of product images and attributes in the catalogue.
Is in-store camera AI legal in the EU?
Many uses are legal, but they must comply with data protection law, and the EU AI Act prohibits certain practices from 2 February 2025, such as emotion recognition in workplaces. Systems that focus on products rather than identifiable people carry far lower legal risk.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
