Research librarySeeing Every Defect
Chapter 03 of 14 · The quality gap and the case for vision first · Draft web edition
A useful inspection dictionary describes what the camera can see, what the inspector must confirm and which production decision follows.
The cloth is woven from this chapter's 10 exhibits: each weft is one exhibit's own data, normalised; each warp group is a section. The probe reads the exhibit's own value where it crosses. Procedural visualisation, not a measurement.
Read-out
Derived from S25 Table 3, fifteen daily values per category, printed page 239. Dashes treated as zero. Inputs: {"Broken thread": [2, 4, 2, 3, 0, 1, 3, 3, 3, 2, 1, 4, 6, 6, 6], "Puckering": [2, 3, 3, 1, 4, 4, 2, 2, 1, 3, 2, 3, 3, 3, 3], "Needle hole": [2, 3, 0, 2, 1, 0, 0, 3, 2, 6, 5, 2, 2, 3, 4], "Open seam": [2, 3, 2, 2, 3, 2, 1, 3, 2, 3, 0, 0, 3, 2, 3]}. Sum each row: 46, 39, 35, 31. Selected categories only; not a complete Pareto or unique-garment count. External study, not a GFT result.
Unit: category records over the study observation period. Status: derived. Sources: S25.
Overture · 1 min
A factory should commission its inspection system against a defect dictionary, rather than an undifferentiated promise to detect faulty cloth. This chapter proposes an illustrative catalogue of 42 operational classes across yarn preparation, weaving, knitting, wet processing, cutting and sewing. The grouping is GFT's design proposal, not a universal classification or evidence that any detector recognises every class. Optical signatures are inspection hypotheses to test on the intended fabric. Several classes can look alike, and the same class can change appearance with fabric, lighting and presentation. Keep observed appearance, suspected process origin, severity and release decision as separate fields.
This chapter proposes an illustrative catalogue of 42 operational classes across yarn preparation
illustrative
Published work provides bounded examples, not a complete factory specification. Deep-GD includes a defect-free category in its reported class list. That nominal category is a reference condition, not another defect. The airbag seam study addresses particular stitch configurations, while the warp-knitting study addresses broken-end inspection. Their vocabularies help define the inspection task; their results do not establish interchangeable coverage across products.[S18][S22][S15]
The catalogue below is deliberately operational. Each code should become a controlled definition, an image collection, a disagreement rule and an escalation route before automatic release is considered. The proposed visual checks and routing rules throughout this chapter are GFT engineering recommendations. They are not measured detection results. The chapter establishes what must be distinguished; later chapters address optics, models, garment presentation and corrective-action workflows.
3.1
Exhibit 097 paragraphs step it
Read-out
| Y01 Nep-like speck | Compact attached fleck | Check attachment |
|---|---|---|
| Y02 Local slub | Short local bulge | Compare nearby yarn |
| Y03 Thick place | Sustained broad segment | Check contextual length |
| Y04 Thin place | Sustained narrow segment | Compare construction |
| Y05 Hairiness | Peripheral fibre fringe | Compare approved finish |
| Y06 Foreign inclusion | Distinct embedded feature | Confirm material identity |
| Y07 Prominent knot | Local junction or nub | Apply agreed sample rule |
Illustrative GFT dictionary. Seven entries per family, six families, 42 entries in total. Candidate cues and checks are design hypotheses requiring local validation, not measured signatures, universal terminology or fixed acceptance rules. Codes are class labels, not source identifiers.
Unit: proposed operational classes. Status: illustrative. Illustrative figures are teaching examples, not measured factory results.
Start the dictionary upstream of the camera's installation point. A mark first seen on incoming cloth need not have originated in weaving. Conversely, an irregular yarn appearance is not sufficient evidence that a spinner supplied defective yarn. Record the observation as an appearance class until a retained sample or process investigation supports attribution. This distinction is especially useful when a factory buys finished cloth rather than controlling yarn preparation itself. Its quality team can document a reproducible complaint without claiming access to the supplier's machinery records.
For the illustrative yarn family, retain separate entries for nep-like specks, local slubs, extended thick places, thin places, excessive hairiness, foreign-fibre inclusions and prominent knots. Their proposed boundaries distinguish compact particles, short bulges, sustained width changes, filament-like fringes and attached junctions. These are working annotation distinctions. They must be reconciled with the purchased yarn and fabric specification before production use. Deep-GD's use of hairiness and lumpy appearances supports distinguishing texture abnormalities from stains, but does not validate this complete family.[S18]
The proposed nep-like class should request a close view of an attached compact fleck and a wider view of its surrounding texture. Keep loose surface debris in a separate review state until attachment is checked. For the slub class, request the visible bulge and enough neighbouring yarn to identify its direction. For thick and thin places, request a longer contextual crop than for an isolated particle. This keeps the annotation tied to a spatial extent rather than to an operator's impression that the cloth looks uneven.
Treat hairiness as a comparison against an approved normal sample from the same construction and finishing condition. A silhouette against a contrasting background may be a useful commissioning view, but qualification must establish whether it predicts the buyer's assessment. Likewise, the proposed foreign-fibre code should describe a visible inclusion without claiming chemical identification. A camera flag can request investigation; it should not certify fibre composition from appearance alone. For a knot, record whether the junction protrudes, affects an intended visible panel or falls within an agreed acceptable appearance.
The illustrative dictionary in Exhibit 9 assigns stable yarn-family codes and candidate optical cues. It is a commissioning worksheet, not a prevalence table. An entry's presence means that the inspection designer must decide whether it applies to the product. It does not mean the factory has observed that defect, or that the proposed view can reliably separate it from every lookalike.
For procurement, require the supplier and factory to agree how short local bulges are distinguished from sustained thick places. Store examples on both sides of that boundary. If the distinction cannot be applied consistently, keep a broader confirmed irregularity code and retain the finer annotation as provisional. Forcing certainty at intake creates an apparently precise supplier scorecard whose differences may reflect inspection language rather than material quality.
3.2
A directional definition must travel with the cloth. The ASTM terminology preview distinguishes a warp-direction void from a filling-direction discontinuity. Those material directions remain meaningful when an image is rotated. A screen label such as horizontal does not. Record warp and weft orientation in the image metadata and retain a reference showing how the fabric was presented. The preview supports these terminology distinctions; it is not a complete acceptance specification.[S0301]
The illustrative weaving family separates broken ends, missing or broken picks, double ends, double picks, reed-spacing marks, unintended floats and snarled yarn inclusions. These are proposed operational labels. The candidate distinctions are absent material, additional material, spacing irregularity, exposed unbound length and a local tangled shape. Deep-GD uses double-end, snarl and miss labels, but a factory must inspect its actual label definitions before mapping those names into the local dictionary.[S18]
For a broken-end candidate, ask whether the narrow feature follows the recorded warp direction and remains distinguishable from an intentional stripe. For a missing-pick candidate, ask the corresponding question across the filling direction. A dark line alone should not decide either class. The proposed decision record should describe continuity, neighbouring yarn structure and the inspection view that exposed the suspected gap. If those observations remain inconclusive, preserve a line-anomaly label and refer the sample for confirmation.
Double-end and double-pick candidates need a different reference question: does the local structure contain additional material relative to an approved repeat? Reed-spacing marks should ask about persistent spacing rather than local yarn thickness. Floats should request enough context to distinguish a missed interlacing appearance from the intended weave. Snarls should request a local shape description before the investigator attributes the event to any particular machine condition. These are proposed checks, rather than guaranteed optical discriminators.
The worksheet in Exhibit 10 keeps these questions separate. It should be reviewed with the supplier's construction terminology. A detector may initially issue a broader directional anomaly flag while the quality lead assigns a confirmed class. That staged approach retains useful localisation without pretending that an early training set already resolves every structural distinction.
Read-out
| W01 Broken end | Warp-aligned gap | Check yarn continuity |
|---|---|---|
| W02 Missing/broken pick | Weft-aligned gap | Check filling continuity |
| W03 Double end | Extra warp-like line | Compare intended repeat |
| W04 Double pick | Extra weft-like line | Compare intended repeat |
| W05 Reed-spacing mark | Persistent spacing line | Check repeat geometry |
| W06 Unintended float | Exposed unbound length | Compare intended weave |
| W07 Snarled inclusion | Tangled local shape | Confirm structure locally |
Illustrative GFT dictionary. Seven entries per family, six families, 42 entries in total. Candidate cues and checks are design hypotheses requiring local validation, not measured signatures, universal terminology or fixed acceptance rules. Codes are class labels, not source identifiers.
Unit: proposed operational classes. Status: illustrative. Illustrative figures are teaching examples, not measured factory results.
Define the relationship between a physical defect and an image crop before annotating long lines. A continuous feature may cross several crops or camera fields. Treat those crops as evidence attached to a candidate event, then use the verified material position to decide whether they belong together. A disagreement about event continuity should remain visible in the record. Do not quietly convert a crop count into a count of defective garments or a length of rejected cloth.
For model qualification, require directional cases in both their normal presentation and plausible rotated presentations. This is a proposed evaluation condition, not an instruction to make the model invariant to every rotation. Material orientation may carry useful information. Preserve it explicitly so that a change in cloth presentation can be distinguished from a change in the process. The dictionary should help the optical and model teams state that requirement without selecting their hardware or architecture for them.
3.3
Exhibit 117 paragraphs step it
Read-out
| K01 Dropped loop | Local loop interruption | Inspect surrounding loops |
|---|---|---|
| K02 Needle line | Persistent wale-like line | Compare normal texture |
| K03 Ladder | Extending loop opening | Retain full extent |
| K04 Hole | Area of open material | Confirm under set tension |
| K05 Unwanted tuck | Unexpected loop motif | Check pattern reference |
| K06 Sinker-line candidate | Repeated surface mark | Confirm structural origin |
| K07 Spirality/course skew | Global axis departure | Apply agreed test method |
Illustrative GFT dictionary. Seven entries per family, six families, 42 entries in total. Candidate cues and checks are design hypotheses requiring local validation, not measured signatures, universal terminology or fixed acceptance rules. Codes are class labels, not source identifiers.
Unit: proposed operational classes. Status: illustrative. Illustrative figures are teaching examples, not measured factory results.
A knitting dictionary should begin with the intended structure rather than borrow woven labels indiscriminately. The smart-sensor study reports a warp-knitting application aimed at broken-end defects. Its reported 98 percent detection rate follows six months of operation on that application. This is evidence for that inspected task, not evidence that all knit faults or garment styles have the same detectability.[S15]
The illustrative knit family uses dropped-loop candidates, needle-line candidates, ladders, holes, unwanted tuck loops, sinker-line candidates and spirality or course skew. These labels separate a local interruption, a persistent line, an extending opening, an area opening, an unexpected loop appearance, a repeating surface mark and a global alignment condition. The boundaries are proposed for annotation and require confirmation by a knitting specialist. In particular, a hole records an observed opening; it does not, by itself, identify the mechanism that created it.
For dropped-loop and ladder candidates, request both a detailed crop and a contextual image showing how far the affected structure continues. For a needle-line candidate, retain a normal comparison from the same fabric family. For unwanted tucks, require an approved pattern reference so an intentional feature is not labelled defective. Sinker-line candidates should remain provisional until the visual mark is connected to a confirmed structural or process observation. Naming a machine component in a class must not turn an image prediction into a maintenance diagnosis.
For spirality or course skew, the proposed boundary is larger than a local defect crop. Record the presentation state, reference axis and measurement method. Avoid assigning a precise alignment failure from a loosely positioned garment photograph. If the assessment requires conditioning, relaxation or a specified test procedure, mark the camera observation as a request for that assessment. The intended inspection gate should determine which presentation belongs in the dictionary.
The proposed classification in Exhibit 11 is therefore accompanied by a presentation requirement. A useful commissioning sample should show the accepted state, the suspect state and any manipulation needed for confirmation. Keep an opened hole, a slack fold and a dark printed feature available as deliberate lookalikes. This gives reviewers a practical way to record what the installed view can and cannot decide.
Create an observability status for every inspection event. Use it to distinguish an adequate view from an obscured, folded or otherwise unsuitable view. An uninspectable area should not silently inherit an acceptable label. Route it for another presentation or an explicit manual check. Record the reason so that repeated presentation failures can be addressed independently of model errors. This is a proposed operating rule, and it should be tested with the people who will handle the material.
The class list should remain stable when presentation improves. What changes is the evidence available for assigning a class, not necessarily the underlying definition. Retaining that distinction allows the quality lead to compare historical records while the inspection team changes fixtures, handling instructions or camera recipes. If improved presentation exposes a distinction that was previously impossible to adjudicate, introduce the refined code through a documented dictionary revision rather than rewriting old records without explanation.
3.4
The proposed wet-processing family contains dye spots, oil-like stains, soil-like stains, water-mark candidates, shade drift, weak or faded print and print misregistration. Deep-GD explicitly distinguishes soil stain, oil stain and dye spot. Its vocabulary is useful evidence that stain classes can be separated in a research dataset. It does not demonstrate that an ordinary colour image proves the substance responsible for any new factory mark.[S18]
For that reason, prefer oil-like and soil-like while the event is awaiting confirmation. Store the visual class and the confirmed substance, if established, in different fields. A maintenance investigation may later connect a stain to a machine or handling step. That conclusion should be added with its evidence and reviewer, rather than embedded in the original detection as if the camera observed the cause. The quality log then remains useful even when the first explanation is rejected.
For dye spots and water-mark candidates, the proposed annotation should capture the region and its surrounding colour context. For shade drift, request a comparison against the approved reference under the qualified inspection recipe. Preserve exposure settings and any reference-capture procedure used for that decision. Do not give a numerical colour-difference threshold authority simply because the software can calculate it. The acceptance threshold and measurement method belong to the product agreement and require their own approval.
For weak print, define the expected motif and reference condition before asking an inspector to grade missing coverage or reduced density. For misregistration, define which pattern features must align and the permissible relationship between them. The proposed distinction is between an appearance or coverage failure and a positional failure. Keep both labels available when they coexist, while recording the material as a single affected unit for unit-based yield calculations.
The worksheet in Exhibit 12 reflects these different decision boundaries. It intentionally excludes inferred chemical composition, fastness certification and bleaching adequacy from the camera's automatic conclusions. A visible mark may be reason to request a test. It is not, without appropriate evidence, the result of that test. Keep specialist laboratory assessments linked to the event without implying that the image alone supplied them.
Read-out
| D01 Dye spot | Local colour concentration | Compare colour reference |
|---|---|---|
| D02 Oil-like stain | Local tonal or sheen patch | Confirm substance separately |
| D03 Soil-like stain | Local mottled region | Confirm substance separately |
| D04 Water-mark candidate | Diffuse ring or boundary | Check retained sample |
| D05 Shade drift | Broad colour departure | Use qualified colour recipe |
| D06 Weak/faded print | Low motif coverage | Compare approved artwork |
| D07 Print misregistration | Displaced motif features | Check alignment reference |
Illustrative GFT dictionary. Seven entries per family, six families, 42 entries in total. Candidate cues and checks are design hypotheses requiring local validation, not measured signatures, universal terminology or fixed acceptance rules. Codes are class labels, not source identifiers.
Unit: proposed operational classes. Status: illustrative. Illustrative figures are teaching examples, not measured factory results.
A proposed commissioning exercise should include repeated captures of an unchanged approved sample as well as captures of suspect material. Use these to ask whether a changed alert follows the material or the inspection conditions. Document the result as part of recipe qualification. If the system cannot distinguish those alternatives reliably enough for the agreed decision, route the class to manual confirmation and state that limitation in its coverage record.
Keep local prevalence estimates separate from external case studies. A print shop, a dye house and a sewing line have different opportunities to observe and introduce faults. Combining their category frequencies into a synthetic factory Pareto would give a misleading denominator. The appropriate local distribution should come from the factory's own confirmed events, with its inspection boundary, product mix and coverage stated. External evidence can suggest classes to investigate; it should not fill missing local counts.
Interlude · the cloth half-woven
Woven: sections 3.0 to 3.4Still to weave: 3.5 to 3.7
Read-out
S24 reports SEM-PLS associations in one Indonesian garment company. S25 reports the share of two dominant categories among 16 jacket defect types. S26 reports defective printed T-shirts in another case. These results are not Cameroonian benchmarks or GFT operating outcomes.
Unit: reported study results. Status: sourced. Sources: S24, S25, S26.
Read this exhibit in its chapter ↗3.5
For the illustrative cutting family, use panel-outline distortion, missing edge material, misplaced or excessive notches, drilling-position errors, frayed or ragged cut edges, fused plies and reversed nap or directional motif. These are proposed operational categories. Their candidate checks compare an observed panel or bundle with the approved pattern, orientation and handling requirements. They should not be represented as a classifier validated by a fabric-surface research dataset.
The training material on cutting distinguishes outline accuracy, edge condition, material faults within cut parts and shade differences between bundled pieces. It also treats fusion as a cutting concern. This supports keeping dimensional and handling checks alongside surface inspection, rather than assuming that a clean fabric image proves an acceptable cut component.[S0302]
For panel-outline distortion, request the approved pattern revision and a qualified presentation before measuring deviation. For missing edge material, identify the expected boundary and whether the relevant edge is actually visible. An apparent omission caused by an overlapping panel should remain a presentation failure until separated. For notches, record the expected position and permitted geometry. For drilling, request the intended location and whether the mark is supposed to penetrate the inspected stack or only the component presented.
For frayed edges, retain a detailed crop alongside the panel identity. For fused plies, require a handling confirmation or another qualified observation if the visible top face cannot resolve separation. For reversed nap or directional motifs, compare the component with the approved orientation rule and its paired components. These are proposed inspection tasks. No universal dimension or tolerance is supplied here, because the dictionary should carry the reference to the actual product requirement.
The worksheet in Exhibit 13 also helps define the inspection unit. A top panel, a bundle and a complete garment are different units. The proposed record must say which was inspected and which material the decision covers. An alert from a visible outer component should not be treated as evidence that every concealed ply was checked. Explicit coverage makes a narrow but reliable inspection more useful than an ambiguous whole-bundle pass.
Read-out
| C01 Outline distortion | Contour departure | Use approved pattern |
|---|---|---|
| C02 Missing edge material | Incomplete contour | Expose full panel edge |
| C03 Notch error | Wrong location or depth | Compare pattern geometry |
| C04 Drilling error | Wrong mark position | Confirm intended penetration |
| C05 Frayed/ragged edge | Irregular cut boundary | Inspect detailed edge |
| C06 Fused plies | Failure of separation | Require handling check |
| C07 Reversed nap/motif | Directional mismatch | Compare paired components |
Illustrative GFT dictionary. Seven entries per family, six families, 42 entries in total. Candidate cues and checks are design hypotheses requiring local validation, not measured signatures, universal terminology or fixed acceptance rules. Codes are class labels, not source identifiers.
Unit: proposed operational classes. Status: illustrative. Illustrative figures are teaching examples, not measured factory results.
Retain the connection between roll position and cut component when it is available. A pre-existing hole discovered after cutting should remain a material observation linked to the earlier roll, rather than being assigned to cutting merely because that is where it was noticed. Conversely, a panel-outline failure should not become a supplier material defect without evidence. Store discovery station separately from suspected origin and from confirmed origin. The same principle applies to stains first found during sewing or finishing.
For rework, keep the original event and add the attempted repair and reinspection result. A replaced panel should retain its relationship to the garment or bundle so the quality lead can identify what actually passed. Closing an event should mean the prescribed action was completed and verified, not merely that someone moved the component away from the inspection station. This proposed record structure supports the later corrective-action chapter without prescribing a particular factory software implementation.
3.6
The illustrative sewing family separates broken stitching, skipped stitches, open seams, seam slippage, untrimmed thread, needle damage and seam puckering. The airbag paper describes skipped, missed and superimposed seam defects for specific stitch configurations. Pan and colleagues describe image-based seamline localisation and profile analysis for puckering assessment. These studies support different inspection tasks; neither establishes universal acceptance limits for the proposed sewing family.[S22][S21]
For broken stitching, the proposed question is whether the required thread path has a confirmed interruption. For skipped stitches, compare the observed formation with the expected stitch sequence. For an open seam, inspect the intended join rather than inferring closure from nearby decorative stitching. Seam slippage should request the agreed confirmation procedure and reference condition. If that procedure is not represented in the camera view, retain a suspect opening class until the required check is made.
For untrimmed thread, record the intended visible zone and the applicable finishing rule. For needle damage, distinguish a visible damaged area from a normal penetration appearance through agreed samples and, where necessary, closer examination. For puckering, define the evaluation presentation and reference grades before collecting training labels. An image class should not imply that the inspector has proved a particular tension, feed or shrinkage mechanism. Store any such diagnosis after the relevant investigation.
The proposed worksheet in Exhibit 14 makes these boundaries inspectable. Keep construction failures, surface appearance and presentation failures separate even if they share a review queue. A garment may have both an interrupted stitch path and an uneven seam appearance. Permit multiple confirmed observations while preventing them from creating duplicate records of the same affected unit.
Read-out
| S01 Broken stitching | Interrupted thread path | Confirm required stitch path |
|---|---|---|
| S02 Skipped stitch | Missing expected formation | Compare stitch sequence |
| S03 Open seam | Unjoined intended join | Inspect intended seam |
| S04 Seam slippage | Suspect opening at join | Apply confirmation procedure |
| S05 Untrimmed thread | Free projecting tail | Apply zone finishing rule |
| S06 Needle damage | Suspect penetration damage | Compare approved samples |
| S07 Seam puckering | Uneven seam appearance | Use agreed presentation |
Illustrative GFT dictionary. Seven entries per family, six families, 42 entries in total. Candidate cues and checks are design hypotheses requiring local validation, not measured signatures, universal terminology or fixed acceptance rules. Codes are class labels, not source identifiers.
Unit: proposed operational classes. Status: illustrative. Illustrative figures are teaching examples, not measured factory results.
A useful sourced comparison is available from the jacket study. Summing its daily entries gives 46 broken-thread records, 39 puckering records, 35 needle-hole records and 31 open-seam records. Exhibit 15 shows only those selected categories, not a complete Pareto. The counts describe that study, not a Cameroon baseline, and should be treated as category records rather than independently verified unique defective garments.[S25]
The practical lesson is to preserve the class boundaries before ranking them. A combined sewing-defect code cannot distinguish the next investigation for an interrupted thread path from the next investigation for seam appearance. However, the relative frequency in an external jacket study should not decide the priority of a new uniform or knitwear line. Start with the product's requirements, then build a confirmed local distribution with documented inspection coverage.
Do not equate visible severity with risk without the product context. A small observed opening at a load-bearing join may require a different disposition from a conspicuous cosmetic mark in a concealed zone. The quality lead should approve the severity rule against the intended use and buyer specification. Keep the class stable while allowing severity and disposition to vary by zone and product. A defect dictionary is a language for decisions, not a substitute for the product's acceptance agreement.
3.7
Point anomaly
Point anomaly: Plain cloth 1, Twill 2, Printed motif 3, Presented garment 3 (illustrative ordinal effort score) · illustrative
Implement the proposed dictionary as a versioned reference used by both the inspection application and the Performance-Management System.
Directional line
Directional line: Plain cloth 1, Twill 2, Printed motif 3, Presented garment 3 (illustrative ordinal effort score) · illustrative
Colour region
Colour region: Plain cloth 2, Twill 2, Printed motif 3, Presented garment 3 (illustrative ordinal effort score) · illustrative
Seam appearance
Seam appearance: Plain cloth 3, Twill 3, Printed motif 3, Presented garment 2 (illustrative ordinal effort score) · illustrative
The method behind it: Worked example: observations and affected garments have different denominators.
GFT planning assumptions: 1 = simple reference task, 2 = added discrimination required, 3 = strong reference/presentation dependence. Scores are assumed priorities, not empirical difficulty, probability, accuracy or interval-scale measurements. Do not average them.
Unit: illustrative ordinal effort score. Status: illustrative. Illustrative figures are teaching examples, not measured factory results.
3.7 · the bodyExhibit 16 above; its figures are illustrative where marked.
Implement the proposed dictionary as a versioned reference used by both the inspection application and the Performance-Management System. Each class should have its stable code, preferred name, accepted synonyms, inclusion criteria, exclusions, required view, measurement boundary and reference images. Keep the production-stage family as an organising field. Do not make it compulsory evidence of origin. Every event should retain the dictionary version under which it was first labelled.
For the proposed event record, retain material or garment identity, discovery station, timestamp, image reference, material coordinates where available, presentation state, candidate classes, confirmed class, reviewer, severity, disposition and repair status. Record model version and confidence separately from the confirmed judgement. Confidence expresses the model's output under its configured interpretation; it should not be relabelled as a physical probability of product failure without validation for that purpose.
Separate unknown appearance from uninspectable presentation. Unknown appearance means usable evidence exists but the approved dictionary does not yet resolve it. Uninspectable means the evidence is insufficient for the required check. Give both a review route. Keep confirmed acceptable examples as carefully as confirmed defects so the training record retains the intended boundary. Borderline cases should keep their adjudication notes, particularly when a new product or finish changes what counts as normal.
Exhibit 16 is an illustrative prioritisation aid for qualification. Its ordinal scores are assumed effort categories, not measured algorithm accuracy. It asks reviewers to consider a feature against different backgrounds and with different evidence needs. A higher score requests a stronger qualification plan; it does not establish that the class is intrinsically harder under every camera or product configuration.
The synthetic profiles in Exhibit 17 serve a narrower purpose. They show possible local contrast shapes for training discussions. They are constructed arrays, not captured images, measured grey levels or Fourier spectra. Their equal sampling positions have no assigned physical pitch. Use them to explain why a crop, a profile and a defect identity are different objects, then replace them with qualified local examples when commissioning.
The proposed review sequence in Exhibit 18 preserves that distinction. Capture and provisional classification precede adjudication. Acceptance follows the applicable rule, while unresolved evidence goes to another check. Automatic model output should not overwrite the reviewer decision. A correction should append its author and reason so later analysis can reconstruct the original evidence and the decision that followed.
For severity, use approved descriptive categories rather than pretending that arbitrary weights have a universal meaning. If a factory later needs a weighted prioritisation score, publish its scale and retain the unweighted event counts beside it. Never report a weighted sum as defective garments. The proposed taxonomy therefore contains no default critical, major or minor assignment that could silently overrule a product-specific decision.
Before releasing a dictionary version, ask the quality lead and process specialists to annotate retained samples independently, discuss disagreements and approve the revised boundaries. Keep a separate challenge set for classes that matter to the product. Record whether each class is automatically classified, automatically flagged for review, manually inspected or outside the station's scope. A class without a qualified detection method remains a coverage gap, even if its name appears in a training menu.
The acceptance artefact should be a class-by-class coverage record attached to the relevant product and inspection recipe. Review it when construction, colour, finish, presentation or acceptance requirements change. Retain the old mapping so a changed class definition is not misread as a changed production trend. Warp and Signal becomes practical at this point: measurement follows a material identity, while the decision retains its evidence and its limits.
Check the record
Every knot is a source this chapter cites; select one to read its record, or filter the register by grade. Publication dates and inherited-source limits stay explicit. Read the paper's full evidence summary.
7 sources · 19 citation marks on this pageMost cited: S18, 5 times
Grade AGrade BGrade CKnot size: citations on this page
All 7 records shown
c03-S15 · Grade B
Online Fabric Defect Inspection Using Smart Visual Sensors
This paper proposes a novel automatic inspection scheme for the warp knitting machine using smart visual sensors.
2013-04-09. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c03-S18 · Grade B
Deep-GD: Deep Learning based Automatic Garment Defect Detection and Type Classification
It is challenging to detect garment defects automatically because of the complexity of images and variety of patterns in textiles.
2024-02-05. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c03-S21 · Grade B
Image analysis for seam-puckering evaluation
The seamline was located with edge detection and Hough transform techniques.
2016-10-14. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c03-S22 · Grade B
Defect Detection Techniques for Airbag Production Sewing Stages
The paper presents a framework for the automatic detection of defects occurring during the airbag sewing stage.
2015-10-25. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c03-S25 · Grade B
Quality control of garment product using DMAIC Six Sigma
The defects were caused by human, machine, material, method, and environmental factors.
2024-06-30. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c03-S0301 · Grade A
Standard Terminology Relating to Fabric Defects, ASTM D3990-12(2020), preview
This terminology covers defects in both woven and knit fabrics.
2020-07. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c03-S0302 · Grade A
Surface Ornamentation Techniques (Embroidery), Trade Theory, revised syllabus 2022, first edition February 2023
Cutting quality is a prerequisite for quality in a finished product.
2023-02. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗S18 International Journal of Electrical and Electronics Research. 2024-02-05.
S22 arXiv. 2015-10-25.
S15 Sensors (MDPI). 2013-04-09.
S0301 ASTM International. 2020-07.
S0302 National Instructional Media Institute, Directorate General of Training, Government of India. 2023-02.
S21 Textile Research Journal. 2016-10-14.
S25 OPSI. 2024-06-30.
Keep the thread · chapter 03 of 14
Move the shuttle along the rack to preview any chapter: its sections, exhibits and reading time. Draft web edition. This research is not investment, legal or tax advice; illustrative models are not measured factory results.
The paper's rack · 14 boltsBolt length: reading time
01Part I: The Quality Gap and the Case for Vision First
When a new factory has no sales history and no legacy data spine, pre-trained computer vision can begin inspection immediately, with payback dependent on verified waste and rework savings.
02Part I: The Quality Gap and the Case for Vision First
Understanding why human eyes miss defects on moving fabric explains why adding more manual inspectors cannot solve a factory's quality problem.
03The quality gap and the case for vision first
A useful inspection dictionary describes what the camera can see, what the inspector must confirm and which production decision follows.
04Part II: Optical Physics, Imaging Hardware and Deep-Learning Architectures
No deep-learning architecture can recover a sub-millimetre broken thread that was blurred by motion or washed out by diffuse ambient factory light.
05Part II: Optical Physics, Imaging Hardware and Deep-Learning Architectures
From classical Gabor wavelets to YOLOv11 and Mask R-CNN, architectural choices trade off sub-millimetre recall against edge inference latency.
06Part II: Optical Physics, Imaging Hardware and Deep-Learning Architectures
On day one, a new Cameroonian line has rolls of good SODECOTON cloth but no local catalogue of defects; one-class learning turns nominal fabric into a detector immediately.
07Part II: Optical Physics, Imaging Hardware and Deep-Learning Architectures
Inspect the defects created by assembly as well as those inherited from cloth. The reported 0.45-second result concerns image processing, while garment handling and dimensional acceptance require separate proof.[^S19]
08Part III: Edge Engineering, Weak-Grid Resilience and Buy-Before-Build Economics
In Cameroon's Douala-Dibamba corridor or northern Garoua belt, a cloud-dependent camera is a single point of failure; edge inference makes vision self-contained.
09Part III: Edge Engineering, Weak-Grid Resilience and Buy-Before-Build Economics
A greenfield garment maker should license proven vision models and own its captured defect dataset rather than hire a bespoke machine-learning research team.
10Part III: Edge Engineering, Weak-Grid Resilience and Buy-Before-Build Economics
Low nominal wages weaken the case for labour substitution. The investable return comes from material yield, avoided rework and protected capacity, each measured without counting the same garment twice.
11Part III: Edge Engineering, Weak-Grid Resilience and Buy-Before-Build Economics
Counting defects without fixing their root cause merely automates scrap accounting; closed-loop integration wires edge cameras directly into the Performance-Management System.
12Part IV: Workforce Transition, Traceability Governance and the Phased Deployment Roadmap
A greenfield Cameroonian factory faces a thin supervisor pipeline; human-in-the-loop vision turns scarce quality leads into high-impact process coaches.
13Part IV: Workforce Transition, Traceability Governance and the Phased Deployment Roadmap
When vision inspection is tied to cotton lot numbers, quality control generates the process and pre-consumer waste evidence that can later support buyer and Digital Product Passport requirements.
14Part IV: Workforce Transition, Traceability Governance and the Phased Deployment Roadmap
Synchronising vision deployment with the eighteen-month Performance-Management System rollout takes a greenfield line from manual logs to closed-loop quality without operational shock.