Can a Camera Catch Fabric Defects Before a Whole Roll Is Wasted?
What Italian textile manufacturers should test before using computer vision for fabric inspection, from lighting and labels to operator review.
Textile manufacturers and quality managers
6 min read
2 November 2026
3 sources
Image
Camera + lighting
→Label
Defect definition
→Test
Separate runs
→Alert
Location + image
→Review
Operator decision
Inspect missed defects and false alarms under real production conditions.
An inspection model sees the images your camera gives it. If the lighting hides a defect, a better model may not be the first fix.
- 01Test optics and lighting before committing to a model.
- 02Collect representative defects across materials and shifts.
- 03Measure missed defects, false alarms and operator workload together.
A small defect can travel a long way
A knitting line produces a flaw that is obvious under inspection lighting but hard to see on a moving roll. If it is found after finishing or cutting, the team may need to trace a larger batch and decide what can be reworked. A camera near the line could alert an operator earlier, provided it sees the defect reliably at production speed.
A European Digital Innovation Hub account of Czech machine-vision company 24 Vision describes a related challenge: moving towards AI inspection required both representative images and a validated imaging setup. It is evidence of one company's development work, not proof that the same model will work on an Italian textile line.
Get the image right first
Choose a specific defect and product family. Test illumination, camera position, line speed and surface variation with the quality team. Glossy fabric, dark dyes, changing yarn and motion can alter what a camera sees. Retain the images and the physical inspection decision so disagreements can be examined.
Agree what counts as an actionable defect before labelling. A mark that is unacceptable on a premium garment may be acceptable on an internal seam. Labels should reflect the actual customer specification and include borderline cases, not only perfect examples.
Give operators a useful alert
Run the model in shadow mode first: record its alerts alongside the existing inspection process without changing the production line. The alert should show the image, location and time, with an operator able to mark it useful, false or uncertain.
Do not let a language model decide whether machinery stops. Any automated line intervention needs a separately engineered safety and control design. For an initial pilot, an alert that helps a trained inspector find a questionable area is a narrower and more measurable goal.
Test what happens when the fabric changes
Split evaluation by production run or time period, not random frames from the same roll appearing in both training and test sets. Include different materials, lighting conditions, shifts and defect types. A model that performs well on one camera and fabric may fail after a change in dye or setup.
Track missed defects, false alarms per hour, inspection time and the amount of material checked or held. Compare outcomes with the existing inspection process and record any change in line speed or staffing before attributing savings to the model.
Use this before approving the work.
- 01Choose one defect and product family.
- 02Validate camera, lighting and line speed.
- 03Keep operator decisions and image evidence together.
- 04Test on separate runs and monitor missed defects.
Research and references supporting this article. Interpretation and recommendations are Thalenor’s.
The knitting-line situation is constructed; the cited innovation-hub case concerns a Czech company, not an Italian deployment.
Published and reviewed on 2 November 2026. This article provides general technology and operational information, not legal, tax, medical or regulatory advice.