Machine vision inspection has been discussed in the garment trim industry for years, yet few lines actually run it well. The obstacle is not the camera but the sheer variety of trims: in one workshop, a woven label means tiny characters at high density, a hang tag means large solid-colour areas, and a poly bag means a continuous transparent web. Lighting, lens and decision logic differ completely between them. Buying one generic system and expecting it to adapt on its own almost always fails.

This guide covers what vision can and cannot check, how to set thresholds for the two hard cases — woven-label characters and hang-tag colour — what to watch in inline poly-bag inspection, and what order volume justifies the equipment.

What vision can and cannot check

Defect type Current practice Vision feasibility
Missing strokes, blurred characters or skipped picks on a woven label Loupe and eye, 10 to 20 percent sampling Feasible, but camera position and lighting must be fixed first
Registration drift and solid-colour deviation on hang tags First-off against a colour book, then experience Feasible, judged on a ΔE threshold
Unsealed seams, trapped air, missing print on poly bags Hand tear tests; whole rolls are often scrapped Feasible, well suited to inline inspection
Out-of-tolerance dimensions, frayed cut edges Caliper sampling; first-off sets the size Feasible with a locating fixture
Hand-feel, softness, odour, solvent residue Hand and nose, or a third-party lab Not feasible; keep human or lab checks

In one line: vision is good at geometry and colour you can see, and bad at physics and chemistry you have to touch. Confusing the two guarantees a misdirected project.

Woven-label characters: the hardest case

A woven label is woven, not printed: character edges are naturally fuzzy, warp threads drift slightly with tension, and contrast collapses when ground and text colours are close. Three things must be solved first. Use low-angle side lighting so the relief of warp and weft casts shadows and stroke edges appear. Guarantee resolution — below 1.5 mm character height you need at least 20 pixels per millimetre. And compare against a template with tolerance rather than pixel-rigid matching, because tension differs slightly from roll to roll and rigid matching condemns good product.

The acceptance criteria must live in the inspection document, not in an engineer's head: permitted broken-thread length, share of blurred character area, number of skipped picks, count of soil spots. Without these thresholds, vision merely replaces a human argument with a machine argument.

Hang-tag registration and colour: let ΔE speak

Hang tags are large-area solid-colour print, and the two most valuable checks here are registration drift and colour deviation. A workable approach: take the customer-approved first-off as the reference, accept below ΔE00 2, treat 2 to 3 as a warning needing recheck, and reject above 3. The tolerance must be agreed with the customer — fast fashion and luxury brands accept very different ranges.

Two traps: metallic effects such as foil and holographic film, and dark grounds such as black or navy, distort in camera and need their own white-balance and exposure calibration rather than sharing parameters with ordinary four-colour work. And textured stocks such as recycled or cotton paper make readings jump, so measure an averaged area rather than a single point.

Poly bags: inline inspection of a continuous web

Poly bags run as a continuous web, so vision suits continuous defects best: missing print, colour-to-colour drift, foreign matter trapped in the seal, scratches on the body, off-position punch holes. Such systems usually sit at the end of the bag-making machine, run at line speed, and trigger an alarm and stop after a run of consecutive defects, avoiding the scrapping of a whole roll — the cost that hurts bag plants most.

One detail often missed: the bag is transparent or translucent, so front lighting gives almost no contrast. Backlighting, in transmission, reveals foreign matter, bubbles and thickness variation far better — the standard approach in film converting.

Why most trim plants still inspect by hand

Cameras are cheap; a stable decision is expensive. Three practical thresholds: changeover frequency, because every new SKU needs a new template and lighting position and five changes a day eats the benefit; order batch, because below a few thousand pieces per style the capital cost will not amortise; and whether the customer actually rejects, because if defects only cause internal rework and no chargeback, manual inspection is enough.

In practice vision arrives first in three situations: high-volume single-SKU hang tags, continuous-web poly bags, and size or composition labels where character precision is critical. All three share clear standards, stable volume and an expensive failure.

A three-step rollout

  1. Offline sampling: run a two-week human-versus-machine comparison on one machine and log both escapes and false calls. Talking about full inspection without that data is gambling.
  2. Workstation assist: install vision at the critical stations — after woven-label cutting, at the tag inspection table — and let it only prompt at first, with the operator confirming the final call. This is the step that wins staff support.
  3. Inline full inspection: once decisions are stable, connect to the line, define alarm and stop logic, and write the thresholds into the technical agreement with the customer so acceptance has a basis.
One principle: vision does not improve quality, it turns a quality standard into executable thresholds. Settle the standard first, then the equipment; if the standard cannot be agreed, equipment only makes the dispute more expensive.

Related articles

Under pressure to supply inspection data or an approved first-off? Send us both a defect sample and an acceptable sample; we confirm the first-off against the thresholds agreed with your customer and write the criteria into the technical agreement.

Ask about inspection criteria and first-off approval →