Research librarySeeing Every Defect
Chapter 07 of 14 · Part II · Draft web edition
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]
The cloth is woven from this chapter's 4 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
GFT proposed arrangement for a cut-make-trim line. Nodes are inspection responsibilities, not a verified installation or proof of complete surface coverage.
Unit: illustrative process sequence. Status: illustrative. Illustrative figures are teaching examples, not measured factory results.
Overture · 1 min
Computer vision should follow the transformation from cloth into a garment, because an acceptable roll cannot establish that its cut panels, seams and finished dimensions are acceptable. The engineering answer is a sequence of inspection decisions, each tied to the operation that can create or expose a fault. The title's 0.45-second figure is reported image-processing time from a trouser-classification study, rather than demonstrated elapsed time for detecting every stitch, seam and dimensional fault in a complete garment. This chapter proposes how to extend coverage without confusing an image result with a production guarantee.[S19]
7.1
The inherited Cameroon sector record reports approximately 95 percent of cotton fibre exported raw and approximately 5 percent processed locally. Its original publication date was not recorded. Those proportions explain the interest in local transformation; they do not describe the performance of a GFT factory or identify a garment quality system already in operation. The recommendations below concern a proposed cut-make-trim inspection arrangement, with incoming cloth treated as an input rather than evidence that downstream assembly will conform.[R06]
GFT recommends organising the quality plan around what changes physically. Receipt establishes the condition of the material presented for cutting. Cutting establishes panel identity and geometry. Sewing establishes the intended joins. Finishing establishes the appearance and dimensions presented to the buyer. Each release should answer a bounded question about a recognisable workpiece. A single generic label such as good garment would hide which operations and surfaces were actually observed.
Start the design with an inspection coverage sheet. For each product family, record the critical construction features, the view needed to see them, the operation after which they become visible, and the person authorised to decide an exception. Distinguish front, reverse and concealed surfaces in plain language. A checked exterior should never imply a checked interior. If the construction hides an important join, propose an earlier inspection before closure and retain the operation reference in the finished item's record.
The proposed topology in Exhibit 32 links these decisions. It is an illustrative reference arrangement, not an implemented GFT line. Arrows indicate the intended passage of material and information. They do not prove that every fault will be visible, that handling will fit production time, or that a station can safely interrupt a sewing machine.
Treat the topology as a proposal to test with production staff. Walk an actual construction sequence with the quality lead before specifying cameras. Ask where a rejected item can be held without losing its identity, where an operator can present an obscured face, and where a confirmation action would interrupt safe handling. The best location is the earliest useful decision point at which the relevant feature can be presented consistently and a practical response remains available.
GFT also recommends a vocabulary that separates fault, observation and disposition. A dark region is an observation; a stain is a reviewed interpretation; rework or reject is an authorised disposition. Store these separately. Otherwise a model's provisional guess becomes a permanent quality allegation, and the factory loses the ability to distinguish a camera problem from an actual assembly problem when reviewing recurring events.
7.2
Related · chapter 039 paragraphs step it
Read-out
GFT teaching arrays, not measurements or Fourier spectra. Nine equally spaced positions with no physical pitch. Warp gap and slub are transverse cuts, stain and hole are centre cuts, stitch and pucker profiles follow the proposed seam direction. No common physical scale or universal class signature is claimed.
Unit: illustrative signed contrast, arbitrary units. Status: illustrative. Illustrative figures are teaching examples, not measured factory results.
Read this exhibit in its chapter ↗For cutting, GFT proposes carrying the incoming defect map forward as a traceable material record. The proposed record should identify the roll, inspected face, travel direction, distance origin and transverse origin. Cutting staff should confirm the orientation after spreading. A coordinate without that boundary is insufficient: a correct location in an inspection image may refer to a different physical position after rewinding, trimming or turning the cloth. This is a design requirement, not a claim about an installed projection system.
Do not begin with a visually impressive marker overlay. Begin with a physical registration trial. Mark a harmless reference on a sacrificial length, record it at receipt and follow it through the intended material route. At the cutting table, compare its observed position with the transferred position. Repeat after the handling operations the factory actually intends to use. The production engineer should define an acceptable registration discrepancy in relation to the clearance needed around defects, rather than accept an arbitrary pixel error.
Keep the transfer rule explicit. A longitudinal coordinate may require an offset for trimmed material and a reversal when the roll orientation changes. A transverse coordinate may require mirroring when the observed face changes. Stretch, skew and uncertain layer alignment need separate consideration. GFT recommends recording the transformation and its verification status alongside the original map, so an invalid mapping can be withdrawn without erasing the original observation.
An illustrative coordinate example assumes a mapped fault at 12.40 metres from the original origin, 0.30 metres removed before spreading, and unchanged direction. Its proposed table position is 12.40 minus 0.30, or 12.10 metres. An assumed exclusion margin of 0.05 metres would reserve 12.05 to 12.15 metres longitudinally before considering the fault's own extent. These values demonstrate bookkeeping only. They are neither measured registration accuracy nor an instruction to cut production cloth.
GFT recommends making splice treatment explicit in the same trial. A mapped exclusion can protect a marker only if the spreading team knows which layers contain the affected region and how a splice changes usable material. Record the instruction given to the cutter, the material actually set aside and any panels subsequently replaced. Do not record all mapped exclusions as avoided waste. The useful result is the reconciled outcome after cutting, with the excluded material and replacement work visible.
Where a proposed projection device is used, test the displayed outline against physical references across the usable table, including its edges. An accurate central alignment should not stand in for an unchecked outer region. Have the operator confirm that the projected instruction is visible under working illumination and remains associated with the current lay. If the camera or projector moves, withdraw the previous registration approval until the intended verification is repeated.
The panel check should have its own reference, preferably the approved pattern revision and an agreed presentation method. Compare the intended perimeter, not a loosely similar silhouette. Specify whether notches, drill marks, grain orientation and paired-panel identity are in scope. An outline that looks correct can still belong to the wrong size or revision. GFT recommends making identity confirmation a prerequisite for geometric comparison and displaying the revision clearly at the station.
For a proposed multi-layer lay, do not infer that an overhead view verifies concealed panels. Agree whether the method checks the upper surface, selected separated panels or every panel after separation. Record that choice in the inspection scope. If the workflow cannot preserve individual panel identity, limit the release statement to the bundle and sampling boundary actually observed. The purpose is to avoid a precise-looking contour drawing becoming an unsupported claim about hidden material.
The outline's proposed pattern-defect performance figure is not reproduced here because its publisher page could not be opened in this session. No verified result in this chapter establishes automatic splice projection or cut-panel metrology at production speed. Those functions remain proposed engineering tasks, with acceptance dependent on a local coordinate trial, agreed presentation and a comparison against the approved physical pattern.
7.3
Published airbag work demonstrates a framework that follows seam paths and checks stitch colour patterns to identify skipped stitches, missed stitches and superimposed seams in lockstitch and double-thread chainstitch constructions. This is relevant evidence for inspecting structured joins. It is not evidence that an apparel installation inherits the same performance on a different thread, fabric, seam construction or camera view.[S22]
For an apparel pilot, GFT recommends selecting a critical seam whose exit from the sewing operation can be presented without requiring the operator to defeat guards or reposition hands unsafely. Define the observed seam segment relative to the operation and garment identity. Keep the camera trigger, sample association and review image connected. An alert that cannot identify which item and which seam segment it concerns offers little useful control once the workpiece leaves the station.
The proposed image-validity gate should precede defect classification. Check whether the expected region is visible, the image is sufficiently sharp for the selected stitch feature, and the presentation belongs to the approved product recipe. If visibility is inadequate, return an inspection-unavailable state. Do not encode a missing image as no defect. The quality lead should determine whether unavailable coverage requires re-presentation, manual review or a production hold for that particular construction.
The garment-sewing attention study reports improvements in mAP@0.5 of 1.7 percentage points with CBAM, 1.9 with SENet and 3.7 with Coordinate Attention in its comparison. Its proposed YOLOv8-FPCA method also changes the detection head and loss treatment. The ablation in Exhibit 33 preserves a common metric; it does not compare those gains with classification accuracy from unrelated datasets.[S17]
Read-out
Reported attention comparisons in the opened primary article. Values are changes in a detection metric on that study's data, not absolute accuracy or a forecast for a factory.
Unit: percentage-point improvement in mAP@0.5. Status: sourced. Sources: S17.
For procurement, request a reproducible comparison on the factory's held-out seam examples before treating an architectural choice as valuable. Keep the capture conditions, test garments and annotation rules fixed between model candidates. Ask the supplier to provide the complete detection results, including misses and false alarms, rather than a hand-picked image gallery. GFT recommends evaluating the operating threshold against the planned review workload and the consequences of escapes, not choosing it solely to maximise a summary metric.
Use a conservative initial response. Show the suspected location, preserve the item identifier and route the item for confirmation. A repeated alert on the same seam segment should not generate repeated defective-garment counts. Conversely, a garment with several distinct faults should retain each reviewed fault. Agree the event-merging rule with the quality lead and keep the original images available for diagnosing whether merged observations concerned the same physical defect.
A machine-stop function needs a separate machinery assessment and qualified integration. The quality model should not be granted direct control merely because its demonstration runs quickly. For the initial proposed deployment, GFT recommends an operator-visible advisory and controlled hold process. Define who can clear the hold, what evidence is needed, and how a station resumes after a camera obstruction or product change. These are proposed operating controls rather than assertions of achieved reliability.
7.4
The opened seam-puckering abstract describes seamline location using edge detection and a Hough transform, followed by rotation, cropping, projection-profile extraction and LOESS smoothing. Extracted features feed a neural classifier for grading. The study supports a route towards repeatable appearance assessment. Its public abstract does not establish the spectral-analysis method suggested elsewhere in the outline, so that separate method is not attributed to it here.[S21]
GFT recommends treating a puckering station as a measurement procedure, not simply another defect camera. Define the seam length presented, garment support, lighting arrangement, pressing state and permitted operator handling. Retain a reference image of acceptable presentation. If an operator smooths the fabric differently on each pass, an apparent improvement in the score may describe presentation rather than a change in sewing quality. The qualification exercise should deliberately repeat placement as well as image capture.
The proposed grading reference should come from the buyer's approved appearance requirements and a recorded set of boundary examples. Ask reviewers to label the same retained samples independently and reconcile disagreements before training or selecting thresholds. Preserve the disagreement record. A model trained on unresolved labels may reproduce that disagreement while displaying a single confident grade. The practical objective is consistent decisions against an agreed requirement, with borderline cases still available for human assessment.
Avoid declaring a cause from appearance alone. For a confirmed puckering event, the mechanic should inspect the relevant sewing settings and materials under the plant's approved troubleshooting procedure. The camera record can show the location and recurrence of the appearance; it does not itself prove that thread tension, feed setting or needle condition caused it. GFT recommends separating the observed symptom from the mechanic's diagnosis and from the corrective action tried.
For the pilot, retain before-action and after-action samples under the same presentation procedure. Compare reviewed outcomes and repeat the placement check. If the appearance changes only when the operator handles the item differently, do not record a process improvement. If the correction holds across independently presented items, document the scope of that evidence. This creates a useful engineering record without turning a local adjustment into a general claim about all fabrics or seams.
Interlude · the cloth half-woven
Woven: sections 7.0 to 7.4Still to weave: 7.5 to 7.7
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 ↗7.5
Untrimmed threads
Untrimmed threads: 96.7 percent, reported per-class performance · sourced
Deep-GD reports 97.16 percent overall classification accuracy across 12 categories using HKBU imagery, bilateral filtering, a Squeeze-and-Excitation network and a Bayesian-optimised Random Decision Forest.
Open seams
Open seams: 90 percent, reported per-class performance · sourced
Defect-free
Defect-free: 86.7 percent, reported per-class performance · sourced
Broken stitching
Broken stitching: 83.3 percent, reported per-class performance · sourced
Seam slippage
Seam slippage: 80 percent, reported per-class performance · sourced
The method behind it: Worked example: image latency leaves handling time unresolved.
Publisher abstract for the Addis Ababa trouser study. Class denominators and confusion matrix were not available in the opened abstract. These values are not a factory acceptance guarantee.
Unit: percent, reported per-class performance. Status: sourced. Sources: S19.
7.5 · the bodyExhibit 34 above; its figures are illustrative where marked.
Deep-GD reports 97.16 percent overall classification accuracy across 12 categories using HKBU imagery, bilateral filtering, a Squeeze-and-Excitation network and a Bayesian-optimised Random Decision Forest. Its categories include fabric-texture and stain conditions. That result does not establish complete inspection of assembled garment interiors or dimensional conformity. Use it as an example of a classification approach, with its dataset boundary retained.[S18]
The Ethiopian trouser study reports 1,000 collected images, a VGGNet-19 feature extractor with an artificial neural network, and 86.7 percent held-out accuracy. The reported processing time is 0.45 seconds per sample. The per-class percentages in Exhibit 34 are reproduced as the abstract's reported performance; their labels are not silently converted into defect recall, precision or customer escape probability.[S19]
The public abstract does not expose the complete confusion matrix, device specification or elapsed handling cycle. GFT therefore recommends obtaining those details before relying on this evidence for equipment sizing. Keep photographs of the same physical garment and closely related production bursts together when partitioning a local trial. This proposed partition rule reduces the chance that the evaluation rewards recognition of repeated material or presentation rather than successful inspection of unseen production.
Before collecting a local evaluation set, GFT recommends agreeing how disputed examples will be decided. Ask reviewers to inspect the physical item where it remains available, because a photograph may omit the very feature being disputed. Retain an uncertain label when a reliable decision cannot be made. An unresolved example can test the review process without being treated as a confirmed positive or negative in headline model metrics.
Keep the trial sequence independent of supplier demonstrations. Prepare a sealed evaluation list, record the approved software and threshold, and retain all outputs before reviewing aggregate performance. If a revised model is proposed after inspecting failures, use fresh held-out material for the next acceptance decision. Retain the earlier result as development evidence. This proposed procedure makes improvement visible while preventing repeated tuning on the final evaluation set from masquerading as independent proof.
For the local trial, propose a qualification sheet covering product family, size, colour, surface, defect definition, presentation and reviewer adjudication. Include acceptable items as well as retained faulty examples. Record which classes are represented inadequately and exclude them from automatic release. Select acceptance criteria before examining the final trial results. Otherwise the team can keep adjusting the definition of success until an attractive result appears, without improving the actual release decision.
An illustrative trial ledger might contain 200 confirmed defective items and 800 confirmed acceptable items, with 180 defects flagged and 40 acceptable items flagged. On these assumed counts, recall is 180 divided by 200, or 90 percent. Precision is 180 divided by 220, or approximately 81.8 percent. The acceptable-item false-alarm rate is 40 divided by 800, or 5 percent. These invented counts explain denominators, not expected GFT performance.
An illustrative trial ledger might contain 200 confirmed defective items and 800 confirmed acceptable items
illustrative
GFT recommends reviewing that illustrative ledger as a workload as well as an algorithm result. The review desk would receive 220 flagged items and the nominally clear group would still contain 20 defective items. An automatic-release proposal must confront that residual risk explicitly. Add no-read and missing-image cases to the operating ledger rather than excluding them from the denominator simply because no model score exists.
The worked example defines a serial boundary deliberately. If a supplier proposes overlapping presentation and computation, request a timed demonstration of the actual queue and item association. Count re-presentation, result acknowledgement and physical diversion in the trial. A quick processor offers little practical benefit when unresolved items accumulate faster than the reviewer can handle them. GFT recommends retaining time-stamped events for each transition so the limiting activity can be identified directly.
7.6
Exhibit 3510 paragraphs step it
Read-out
Assumed buyer limits 495 to 505 mm. Each assumed reading has a conservative plus or minus 2 mm bound. A full interval inside the limits qualifies for automatic acceptance under the illustrated rule; overlap requires review.
Unit: illustrative chest width, millimetres. Status: illustrative. Illustrative figures are teaching examples, not measured factory results.
Garment measurement is a separate claim from defect classification. The opened university record describes segmentation with UNet and key-point extraction; it reports average measurement errors of 1.27 centimetres for blazers, 0.747 for dresses and 1.012 for skirts. Those results concern photographed garments and do not validate a proposed calibrated flat-lay station. They also give no basis for promising sub-millimetre repeatability on a production table.[S0701]
GFT proposes a fixed measurement plane, a verified scale reference and an approved garment-presentation procedure for a flat-lay trial. Define each measurement by its physical endpoints and construction state. Body length, chest width and sleeve length need written conventions that the buyer and quality lead understand in the same way. Resolve whether the garment is relaxed, fastened, pressed or otherwise conditioned before comparing camera measurements with the tech pack.
The proposed measurement budget should include scale calibration, lens effects, landmark choice, fabric placement and repeated operator handling. Pixel density alone is insufficient as an acceptance argument. Use a stable reference artefact to examine the optical mapping, then repeated garments to examine the full procedure. Keep these exercises separate in the record: a successful reference check cannot establish repeatability when a soft garment is removed and laid down again.
The illustrative example in Exhibit 35 shows why an apparently in-tolerance reading can still require review. Each interval represents an assumed reading plus or minus the same conservative bound. It is a decision illustration, not a confidence interval derived from observations. No actual buyer specification, station accuracy or GFT production measurement is asserted.
Under that illustrative rule, the reading of 504 millimetres spans 502 to 506 and overlaps the upper buyer limit. Route it for agreed remeasurement rather than automatic acceptance. The reading of 508 spans 506 to 510 and lies wholly outside the assumed limits. That distinction prevents uncertain overlap from being confused with clear evidence of dimensional nonconformity. Neither state should be decided by a sewing-defect classifier's confidence score.
GFT recommends a measurement-system trial in which reviewers repeat placement and measurement without seeing their earlier answers. Compare the camera procedure with the agreed reference method and examine persistent bias separately from scatter. Investigate results by measurement point and product family. An average across easy and difficult landmarks can conceal an unusable point, just as a successful overall defect score can conceal a poorly detected class.
Calibration recovery also needs an explicit proposed procedure. Record the last successful reference check and identify production assessed since that point if the next check fails. The quality lead should decide the affected release boundary from the evidence available, rather than simply recalibrating and erasing the failed state. Preserve the original readings, the failed check and the recovery decision. This makes a calibration failure a traceable quality event with an accountable disposition.
For the operator interface, GFT recommends displaying the measured dimension, applicable limits and review reason together. Keep unavailable, borderline and outside-limit states visibly different. Allow a reviewer to record a repeat measurement without overwriting the initial result. If the revised outcome changes disposition, retain who authorised the change and which method resolved it. The resulting history should explain the decision to a production supervisor without requiring access to model internals.
Release only the dimensions that the qualification actually supports. A contour may support an external width while leaving a folded collar or internal construction unmeasured. Preserve a per-point result and a reason for any unavailable measurement. If the buyer changes the measurement convention or tolerance, review the release rule before reusing the previous recipe. The record should identify the approved specification revision alongside the calibration and presentation versions.
7.7
Vision should support preventive engineering by making repeated symptoms inspectable. GFT recommends connecting confirmed stitch faults to the machine, operation, material recipe and maintenance event. Use the connection to ask a focused question: did the symptom persist after the approved intervention under comparable conditions? Avoid describing correlation with a needle change as proof that wear caused every fault in the cluster.
The proposed maintenance procedure should retain the needle specification, change reason, responsible role and verified restart sample. Follow the machine and needle supplier's applicable instructions when setting replacement intervals and adjustment methods. No universal needle life is specified here. A replacement programme should be justified for the actual materials and operations, with scheduled checks and symptom-triggered review documented as different reasons for intervention.
After a change, require the quality lead to inspect the restart output using the agreed method. Record both a successful restart and a failed restart. A cleared dashboard alarm is not proof that the garment is acceptable, and absence of further camera alerts is insufficient if the view has become obstructed. The proposed restart gate should therefore confirm image availability as well as the reviewed construction feature.
GFT recommends separating maintenance learning from individual performance scoring. Begin with the physical fault, construction and process conditions. Ask operators what changed and retain their explanation as an attributed observation, rather than automatically converting a camera event into blame. A pilot depends on candid reporting of presentation difficulties, borderline examples and workarounds. Those reports are useful engineering inputs even when the model appears confident.
For commissioning, propose distinct sign-offs for visibility, detection, handling and correction. The quality lead owns defect definitions and release scope. Production owns presentation and material movement. Maintenance owns safe mechanical intervention. The digital lead owns event identity, model version and reproducible results. Assign an accountable owner for resolving conflicts between these boundaries. This proposed responsibility split prevents a successful software demonstration from substituting for an accepted factory procedure.
A proposed daily review should reconcile material movement with inspection events. Examine items awaiting confirmation, items held after maintenance and items whose identity was lost during re-presentation. An unresolved hold must not disappear because a new shift signs in or a garment is photographed again. Assign a physical holding location and an accountable role, then close the record only when the reviewed item has a documented disposition.
The handover package should include the approved recipes, retained reference examples, unresolved class gaps, response instructions and restart checks. State where manual review remains necessary and what evidence would permit an expanded scope. Keep power resilience and wider information-system design with their dedicated engineering workstreams; this chapter's acceptance boundary is whether the right garment feature is observed, judged and acted upon in time to support production.
The practical recommendation is to pilot a visible critical seam and a bounded finished-item check, measure the handling burden, and expand only when the evidence supports the next feature. A fast image result is useful within that sequence. It becomes a production control only when identity, coverage, disposition and corrective action are verified together. That is how measurement can remain attached to the material as cloth becomes a garment.
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 · 17 citation marks on this pageMost cited: S19, 5 times
Grade AGrade BGrade CKnot size: citations on this page
All 7 records shown
R06 · Grade C · inherited record
Inherited evidence R06
About 95 percent of cotton fibre exported raw against 5 percent processed locally; industrialisation target of 50 percent local processing.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
c07-S17 · Grade A
Improved YOLOv8 garment sewing defect detection method based on attention mechanism
Construct the dataset with sewing defects as the object of model training.
2024-09-09. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c07-S18 · Grade A
Deep-GD: Deep Learning based Automatic Garment Defect Detection and Type Classification
The proposed Deep-GD model achieves the overall classification accuracy of 97.16%, which is comparatively superior to the existing techniques.
2024-02-05. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c07-S19 · Grade A
Development of an automatic garment defect recognition system using a machine learning approach
This subjective, fatigue-prone process leads to inconsistent defect detection and limits production throughput.
2026-09-06. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c07-S21 · Grade A
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 ↗c07-S22 · Grade A
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 ↗c07-S0701 · Grade A
An Intelligent Solution for Automatic Garment Measurement Using Image Recognition Technologies
UNet as a backbone network has been used for mask retrieval.
2022-04-28. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗S19 SAGE, Textile Research Journal. 2026-09-06.
R06 Cameroon Trade Hub. Publication date unrecorded; inherited source-text check.
S22 arXiv, author-deposited manuscript. 2015-10-25.
S17 JVE International, Journal of Measurements in Engineering. 2024-09-09.
S21 SAGE, Textile Research Journal. 2016-10-14.
S18 FOREX Publication, International Journal of Electrical and Electronics Research. 2024-02-05.
S0701 MDPI, Applied Sciences; author record at Kaunas University of Technology. 2022-04-28.
Keep the thread · chapter 07 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.