Quality & Compliance

Fabric Inspection & Classification

Manual and AI-based defect detection.

Advanced fabric inspection and classification goes beyond a visual scan for holes or stains: it involves applying a structured grading system (such as the widely used point/4-point systems referenced in buyer manuals) consistently across shifts and inspectors, and increasingly integrating AI-based automated defect detection systems that use camera arrays and machine vision to flag defects at line speed. The practitioner's core job is calibrating the balance between inspection speed, sampling depth (100% inspection versus statistical sampling) and defect classification accuracy, since under-inspection lets defective fabric through to costly downstream cutting while over-inspection slows throughput and adds labour cost without a proportional quality gain.

The harder work is defect classification consistency — distinguishing a critical defect (that renders a panel unusable) from a minor one (that may be acceptable depending on garment placement) — and calibrating both human inspectors and machine-vision systems against a shared, documented standard so that grading doesn't vary by inspector, shift or machine vendor. AI-based systems can dramatically increase detection speed and consistency for well-defined defect types (holes, oil stains, weaving faults) but need substantial training data and ongoing recalibration for new fabric constructions, colours or defect types, and still typically require human review for ambiguous or novel defects rather than fully autonomous accept/reject decisions.

How the work is done

1
Define the grading standard
Agree the defect classification system (e.g., 4-point system) and acceptable quality level with the buyer before inspection begins.
2
Calibrate inspectors and/or vision systems
Train human inspectors on defect recognition and, where used, calibrate machine-vision systems against a labelled reference defect set for the specific fabric type.
3
Set sampling plan
Decide 100% inspection versus statistical sampling based on fabric value, defect risk history and the buyer's agreed plan.
4
Run inspection
Pass fabric over an inspection table or machine-vision line at a speed matched to inspector/system detection capability, flagging and logging defects with location and type.
5
Classify and score
Tally defect points per the agreed system and calculate the roll or lot's overall grade against the acceptable quality level.
6
Disposition and feedback
Accept, downgrade, or reject the roll/lot, and feed defect location/type data back to the relevant upstream process (weaving, dyeing, finishing) to address root cause.

Decisions you have to make

100% inspection vs statistical sampling?: 100% inspection catches more defects but costs more labour/time per metre; statistical sampling is faster and cheaper but risks letting defective rolls through — base the choice on fabric value, known defect risk history and the buyer's agreed plan.
Human inspection vs AI-based machine vision, or both?: Machine vision offers speed and consistency for well-characterized defect types but needs training data and struggles with novel or ambiguous defects; many mills run both in combination, with vision systems doing first-pass screening and humans reviewing flagged or borderline cases.
How to classify a borderline defect (critical vs minor)?: Borderline classification should be resolved against the documented standard and, where genuinely ambiguous, escalated for a second opinion rather than left to individual inspector judgement — inconsistent classification erodes trust in the grading system faster than any single wrong call.
How often to recalibrate inspectors and vision systems?: Recalibration should be triggered by changes in fabric construction, colour, or a drift in inter-inspector agreement rates, not left on a fixed calendar schedule alone, since drift can occur well before or after a scheduled date.
Accept a roll with defects clustered in one area vs spread evenly?: Clustered defects may allow the roll to be partially used (cutting around the defect zone) while spread defects may downgrade the whole roll — this decision needs to reference actual marker/cutting yield impact, not just total defect count.

Key metrics (indicative)

Inter-inspector agreement rate

indicative working range, track against baseline

Low agreement between inspectors signals inconsistent grading that undermines the reliability of accept/reject decisions.

Defect detection rate (vision system vs human baseline)

track against baseline for the specific defect type and fabric

A vision system's value depends on matching or exceeding trained-human detection rates for the defects that matter most.

Escape rate (defects found downstream that inspection missed)

indicative working range, track against baseline

Escapes into cutting or sewing are far costlier to correct than catching the same defect at fabric inspection.

First-pass roll acceptance rate against agreed quality level

per the buyer's agreed plan

Tracks overall upstream process quality and the effectiveness of feedback loops to weaving/dyeing/finishing.

Inspection throughput (metres per hour per inspector/line)

track against baseline for the fabric type and inspection method

Throughput must be balanced against detection accuracy; pushing speed without checking accuracy degrades the value of inspection.

Metric targets are indicative working ranges, not standards or legal limits.

Common pitfalls

Relying on statistical sampling for a fabric with a known history of clustered defects — the sampling plan misses the very defects it was least equipped to catch.
Deploying an AI vision system without ongoing recalibration as fabric colour or construction changes across orders — detection accuracy silently degrades and defective fabric passes through undetected.
Allowing inconsistent defect classification between shifts or inspectors without a documented, shared standard — grading decisions become effectively random and erode buyer trust.
Treating inspection as a pass/fail gate without feeding defect data back to the upstream process — the same defect recurs order after order because the root cause was never addressed.
Pushing inspection line speed up to hit throughput targets without validating that detection rate holds at the new speed — defects begin escaping downstream even though inspection volume looks healthy.

Advanced notes and limits

AI-based machine vision inspection has matured well for high-contrast, well-defined defects (holes, oil stains, broken picks) but remains less reliable for subtle defects like slight shade variation or texture irregularities, where human judgement still generally outperforms current commercial systems.
The trade-off between 100% inspection and statistical sampling isn't resolved by simply buying faster equipment — even at high machine-vision throughput, the real constraint often becomes downstream review capacity for flagged defects, which is still human-paced.
Point-based classification systems provide consistency but can obscure the real cost impact of a defect (a critical defect in a highly visible garment panel location matters more than the same defect count would suggest) — advanced practice increasingly links defect location to actual marker/cutting impact rather than treating all points equally.
Cross-mill or cross-vendor consistency in defect classification remains an unresolved practical challenge: even with a shared documented standard, inspector training quality and vision-system calibration vary enough between facilities that a lot graded as acceptable at one site may be classified differently at another.

Worked example

Applying an AQL sampling plan to decide accept/reject on an incoming fabric or finished-garment lot

Lot size: 3,200 units
Inspection level: General Level II
AQL (major defects): 2.5, per the buyer's agreed plan
Sample size code letter for this lot size at Level II: L (per standard AQL sampling tables)
Sample size for code L: 200 units
Accept/reject numbers for AQL 2.5 at n=200: Accept ≤10 defects, Reject ≥11 defects (illustrative, confirm exact figures against the specific table edition in use)

1Confirm lot size (3,200 units) falls in the range mapped to sample size code letter L under the agreed AQL table and inspection level
2Draw the sample size of 200 units randomly across the lot, not concentrated from one carton or production block
3Inspect each sampled unit against the buyer's defect classification (critical/major/minor) and tally defects by category
4Count total major defects found in the 200-unit sample; suppose 12 major defects are found
5Compare 12 against the accept number of 10 for AQL 2.5 at n=200: since 12 exceeds the accept threshold, the lot is rejected on major defects
6Document the specific defect types and frequency driving the rejection so the factory can target root-cause correction rather than a blanket re-inspection

With 12 major defects found against an accept number of 10 in a 200-unit sample, this lot is rejected under the agreed AQL 2.5 plan; the inspector should issue a detailed defect breakdown to the factory to support targeted rework rather than simply returning a pass/fail result.

Case study

Context

A third-party inspection agency conducting final random inspection (FRI) on a woven shirt order began seeing a pattern where lots passed the numeric AQL sampling result but the brand's own retail returns data showed an elevated rate of loose button complaints from the same production window.

Problem

Button attachment strength was included in the inspection checklist as a visual check only (present/absent, visibly secure), without a pull-test measurement, so a marginal but systematic under-tensioning in the button-sewing machine setting was not being caught by sampling even though it was producing buttons that loosened after a few wears.

Action

The inspection agency proposed adding a simple hand pull-test with a defined force threshold to the checklist for button and snap attachments on this and similar programs, and the factory's quality team began spot-checking sewing machine tension settings against a reference gauge at shift start.

Outcome

Retail return rates for loose buttons on subsequent shipments from the same factory dropped notably, and the pull-test check was adopted into the brand's standard final inspection checklist for garments with button or snap closures across other factories in its supplier base.

Audit checklist

Sample drawn randomly across the full lot and multiple cartons/production blocks, not concentrated from one location.
Defect classification (critical/major/minor) applied consistently against the buyer's agreed definitions, not inspector judgement alone.
Sample size and accept/reject numbers taken from the correct AQL table edition and inspection level agreed with the buyer.
Measurement-based checks (pull-test, seam strength, dimensional tolerance) included for defect types that a visual check alone cannot reliably catch.
Defect data recorded by specific type and location, not only as an aggregate pass/fail count, to support root-cause correction.
In-line and end-of-line inspection results cross-checked against post-shipment returns/complaints data to catch systematic gaps in the checklist.
Inspector calibration/training verified current, since defect judgement consistency depends on trained, calibrated inspectors.
Root-cause corrective action tracked and verified on a subsequent lot, not just recorded as closed after the rejected lot is reworked.

Glossary

AQL (Acceptable Quality Limit): A statistical sampling standard defining the maximum percentage of defective units in a sample that can be considered acceptable for a given lot, used to decide accept/reject on a sampled inspection rather than 100% inspection.
Critical defect: A defect that could pose a safety hazard or violate a mandatory requirement, typically subject to the strictest acceptance criteria in an inspection plan.
Major defect: A defect likely to reduce the usability or saleability of the product significantly, though not a safety hazard, such as a broken stitch or visible stain.
Minor defect: A defect unlikely to reduce usability materially but that departs from the specification, such as a small thread end or minor shade variation within tolerance.
Final random inspection (FRI): An inspection performed on finished, packed goods shortly before shipment, sampling randomly across the lot to make an accept/reject decision.
In-line inspection: Inspection performed during production, on the line or at intermediate process stages, intended to catch defects early before they accumulate into a large finished lot.
Sample size code letter: A reference value from an AQL table that maps lot size and inspection level to the number of units that must be sampled for inspection.
Pull test: A measurement-based check applying defined force to a component (button, snap, trim) to verify attachment strength against a minimum threshold.
Defect classification: The agreed system (typically critical/major/minor) for categorizing defect severity, which determines how many occurrences are tolerated under the sampling plan.
Corrective and preventive action (CAPA): The documented process of identifying the root cause of a defect, correcting the immediate issue, and implementing a preventive change to stop recurrence.

Practice questions

1. A lot of 5,000 units is inspected at General Level II with AQL 2.5 (major), sample size 200, accept number 10. The inspector finds 9 major defects. What is the decision, and what should still happen?
2. Why did the button-attachment issue in the case study pass inspection despite being a real, recurring problem?
3. What is the risk of concentrating a sample draw from a single carton or pallet rather than spreading it across the lot?
4. How should an inspector distinguish a major defect from a minor one when the buyer's written definitions are ambiguous for a specific case?
5. A factory reworks a rejected lot and resubmits it for inspection. What should the inspection team verify beyond the numeric pass/fail on the resubmitted lot?
6. Why is cross-checking inspection results against post-shipment retail returns data valuable even when AQL sampling passes are consistently high?

Sub-topics in this chapter

Four-point inspection: Standard manual grading system that scores fabric defects by size to accept or reject rolls.
AI fabric defect detection: Deep-learning models trained on defect images to auto-classify faults on moving fabric.
Camera-based inspection: Line-scan or area cameras with lighting rigs that capture full-width imagery for analysis.
Defect mapping: Digital map of every defect on a roll used to skip faults in the cutting-room marker.
Roll grading: Assigning a quality grade to each roll based on defect density and severity.
Shade grouping: Sorting rolls into shade lots so garments cut from mixed rolls do not show colour variation.

Lessons that teach this chapter

Where this chapter is applied

The value chain stages that use this chapter's skills — chapter to stage to skill.

Study tools for this chapter

Saved on this device only

Study without an account, or sign in to save progress across devices. See the full chapter syllabus.

All chapters and the 42-stage course