Can Computer Vision Spot a Shirt Defect? Start With One Check

Put two photographs of the same black shirt on a desk. In one, the placket lies flat under even light. In the other, a fold hides a button and a bright reflection crosses the fabric.

Before judging the shirt, an inspection system has to make sense of the photographs. That is why a sensible computer-vision project begins with the camera setup and a precise question—not a promise to automate every quality check.

Choose one visible condition

Consider a proposed pilot that helps inspectors flag a missing button on a specified shirt style. Define the expected positions, the required view and the conditions under which the picture is unusable.

Keep other questions outside that first test. Button attachment strength, comfort, shrinkage and the feel of a seam need different evidence. A front photograph should not be asked to certify the entire garment.

TRYBUY.IN’s Black Solid Cotton Shirt for Men is used here as a catalog reference for discussing visible construction. The photograph is not a defect example, and this article does not claim that TRYBUY.IN has deployed an AI inspection system.

Understand what an anomaly model actually contributes

MVTec’s industrial anomaly-detection benchmark includes defect-free training images and test images with and without defects. It illustrates a research approach to identifying departures from expected appearance; it does not establish accuracy for a particular shirt factory. Its published dataset licence also restricts commercial use. MVTec AD dataset and licence information.

A commercial team should therefore use appropriately authorised data and evaluate a system on its own intended inspection conditions. A result on a public benchmark is not a substitute for that work.

Build a small, controlled inspection station

Keep the view repeatable

Specify camera position, working distance, background and lighting. Present the shirt consistently so expected details are visible. Record whether the garment is laid flat or held on a fixture; do not mix methods without testing the effect.

Include a “cannot assess” outcome for obscured or blurred images. An operator should be able to request another photograph rather than being forced to choose pass or fail from poor evidence.

Agree the labels before collecting examples

Ask experienced inspectors to define what counts as the target defect and what remains acceptable. For an embroidery check on a men’s kurta, an intentional motif variation must not automatically be labelled faulty.

When inspectors disagree, resolve the standard first. Training a system on inconsistent labels only moves the disagreement into the software. Keep notes explaining borderline examples and how the final label was chosen.

Separate development from evaluation

Do not place near-identical photographs of one garment on both sides of the test. Hold back different garments and, where practical, a later production batch. Otherwise the evaluation may reward recognition of familiar images rather than reliable inspection.

Count missed defects and unnecessary holds separately

A missed defect can reach the customer. An unnecessary hold can interrupt production and consume an inspector’s time. Report both outcomes, not just a single overall accuracy figure.

For every flagged item, preserve the original image, the model’s suggested region or reason, and the inspector’s decision. Review a sample of passed items too. Looking only at flagged garments cannot reveal what the model overlooked.

Keep the pilot advisory: flag for review, do not automatically reject stock. A useful system should make the inspector’s next action clearer, not bury the team in unexplained alerts.

Recheck whenever the product or setup changes

A new colour, reflective decoration, altered camera or different garment presentation should trigger a review. A successful test on solid shirts does not prove that the same thresholds work on ornate kurtas.

Boys’ garments need their own size and presentation checks. Smaller dimensions can change the apparent scale of details in the image; they should not be treated as adult garments merely placed farther from the camera.

Brief answers

Can a catalog image serve as the inspection standard?

It can illustrate a style, but use controlled, approved inspection references for quality decisions.

Can computer vision replace a physical quality team?

A limited visual pilot should support the team. It cannot establish every property that matters when clothing is worn.

What is a sensible first result?

A measured, repeatable flag for one clearly defined condition, with a workable human-review process.

Discover TRYBUY.IN through the details of the clothes themselves. Technology is useful when it helps people pay closer attention.