Research librarySeeing Every Defect
Chapter 12 of 14 · Part IV · Draft web edition
A greenfield Cameroonian factory faces a thin supervisor pipeline; human-in-the-loop vision turns scarce quality leads into high-impact process coaches.
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
Process synthesis of the STFPM human-in-the-loop inspection workflow and the edge cascade. Adapted as a GFT recommendation for fabric and garment inspection; it is not a claim that this exact workflow is already operating at a GFT facility.
Unit: . Status: sourced. Sources: S34, S07.
Overture · 1 min
The correct deployment unit is not a camera that replaces an inspector. It is a supervised decision cell in which software clears validated nominal work, people resolve exceptions, and supervisors use the resulting evidence to remove causes. This division protects release decisions while the model encounters new cloth, new seams and early production instability. It also gives inspectors a route into higher-value technical work rather than asking them to compete with a machine on continuous watching. The human role must include authority to question the screen, stop the process and improve the workflow. Evidence from industrial anomaly inspection supports exception routing, while International Labour Organization research supports worker involvement in decisions about how AI affects skills and control.[S34][S1201]
12.1
An early garment line is an unstable inspection domain. Fabric construction changes between orders. Dye lots alter the background. Lint, folding and roller vibration change the image. A model can therefore produce a confident score outside the conditions in which its threshold was validated. Confidence is a model output, not proof that the physical item conforms. A greenfield deployment must treat that gap as an operating risk, not as a footnote to a benchmark.
The manual alternative is also insufficient. The inherited textile-inspection evidence places manual defect detection at 60 to 70 percent accuracy, with 20 to 30 percent of defects missed.[R29] A metal-textile inspection study reports a wider manual detection range of 60 to 80 percent and explains why fatigue, repetition and individual variation prevent stable classification.[S34] These figures do not justify swapping one fallible decision-maker for another. They justify an architecture in which each handles the work it can perform most reliably.
That architecture has three decision lanes. Frames inside a validated nominal envelope may receive an automatic PASS. Flagged or ambiguous frames go to an inspector with the original image, the anomaly location and the relevant defect standard. A confirmed critical defect, an unresolved image or a failed station check goes to HOLD and escalates to the quality lead. Automatic rejection of a roll, bundle or garment lot should not be the default. Physical disposal and lot release carry commercial consequences that require an accountable person.
Exhibit 55 makes the decision rights visible before any software is configured.
The loop is deliberately asymmetric. Software may reduce the volume of routine screening, but it may not widen its own authority. A model update changes the screening function only after an approved test. A failed camera check routes work to review rather than allowing the last known model to continue silently. The inspector can disagree with an anomaly map, and that disagreement remains in the record. The supervisor can stop automatic PASS when the product, optics or process moves beyond the validated envelope.
This design also prevents a common interface error. Showing only a red or green result asks the inspector to obey an unexplained score. Showing the cropped region, the physical location and the applicable defect definition gives the inspector something to inspect. The metal-textile study uses anomaly heatmaps for precisely this guidance and reports that unfamiliar visual representations require careful introduction.[S34] Training must therefore cover the meaning and limits of the display, not just which button to press.
Human triage does not guarantee zero escapes. It creates a place to detect, contain and learn from them. Automatic PASS still needs a documented audit sample chosen independently of the model's score. If an audit finds an escape, the response is to hold affected work, review the operating envelope and test the threshold. The person who raises the alarm should be thanked for exposing a system problem, not measured as an obstacle to automation.
12.2
Related · chapter 068 paragraphs step it
Read-out
Illustrative operational trajectory anchored on cited benchmark endpoints: manual baseline recall (65 percent midpoint of 60 to 70 percent, R29); day-one PatchCore/cascade recall (>90 percent, R29, S04, S07); initial recall-prioritised cascade false-positive forwarding rate (49.3 percent in S07) dropping via threshold calibration and hard-negative active learning (29 to 45 percent inference reduction in S07); and automated 42-class specificity rising as local verified defect crops fine-tune YOLOv8/YOLOv11 models (S11, S34).
Unit: percent of inspected frames or defects. Status: illustrative. Sources: R29, S04, S07, S11, S34. Illustrative figures are teaching examples, not measured factory results.
Read this exhibit in its chapter ↗An exception decision becomes useful training material only when its context survives. A one-tap confirmation should store the crop, original frame reference, roll or bundle identifier, station, model version, anomaly score, defect class, severity and inspector disposition. Where the inspector cannot decide, the correct label is UNCERTAIN with escalation, not a forced guess. This discipline turns routine verification into a curated evidence set without pretending that every tap is infallible ground truth.
The first review belongs to the quality lead. Ambiguous crops, new defect forms and disagreements between inspectors should enter a consensus review before they influence a threshold. Duplicate views of the same physical flaw must be grouped so that repeated frames do not dominate the training set. Images from one roll or garment style should not appear in both training and held-out test partitions without that relationship being disclosed. These are basic defences against teaching the model the habits of a shift rather than the physics of the defect.
Active learning should remain offline. The station must not change its own weights after each confirmation. A technician prepares a candidate dataset, the quality lead checks label quality, and the model owner evaluates a candidate version against critical-defect recall, false alarms and subgroup performance. Release requires a named approver and a rollback path. This separation matters because a convenient touchscreen can otherwise become an uncontrolled model-training channel.
Label governance must also preserve the difference between NOT DEFECTIVE and NOT YET UNDERSTOOD. A harmless lint fleck can become a useful negative example only after the inspector establishes that it is removable contamination rather than a stain or foreign fibre. A novel texture should enter an investigation set until the quality lead decides whether it represents acceptable variation, a new defect class or an optics problem. Combining these cases into one nominal label would teach the model to ignore genuine process changes. The review queue therefore needs a small taxonomy of uncertainty as well as a taxonomy of defects.
The edge cascade study shows why better human labels matter operationally. On its disjoint benchmark, the recall-prioritised first stage forwarded 53.4 percent of frames, and 85 percent of those forwarded frames were false alarms. The authors estimate that tighter calibration could reduce inference work by 29 to 45 percent, while explicitly positioning the system as AI-assisted triage rather than autonomous acceptance.[S07] A high false-alarm rate is therefore not merely a computing issue. It determines how often a person is interrupted and how quickly alert fatigue can develop.
The worked example changes the management question. Asking how many frames the model classifies per second ignores the bottleneck experienced by the inspector. The better question is how many consequential interruptions the station creates, and whether each interruption adds information. Review queues should be measured by fabric type, station, model version and shift. A sudden increase may indicate lint, changed lighting or a new print rather than a genuine deterioration in production.
Inspectors should not be rewarded for agreeing with the model. Such an incentive teaches blind acceptance and contaminates the labels. Nor should the number of taps become a productivity target, because it rewards unnecessary forwarding and fast, careless review. Useful measures are audit agreement on independently reviewed samples, correct escalation of critical cases and the proportion of dispositions with complete context. The purpose of active learning is to make human judgement more informative, not to make it more frequent.
12.3
The ergonomic case begins with the task, not with a claim that screens are inherently healthier than fabric. Manual inspection combines sustained attention, repeated eye movement and a body position fixed by the perch or table. A comprehensive fabric-inspection review identifies lack of concentration, human fatigue and time consumption as central drawbacks of manual detection.[S03] Replacing continuous web watching with a poorly designed alarm screen would merely exchange visual fatigue for cognitive overload.
The task should be redesigned around selective attention. An inspector reviews a stationary crop with a visible region of interest, compares it with the physical standard and records a disposition. The inspector no longer needs to search every centimetre of moving cloth for a rare event. The metal-textile study reports an initial workload reduction above 50 percent, with a calculated 69.4 percent reduction when results were scaled, but its authors call for longer production trials and further investigation of worker relief.[S34] Those early results support testing the mechanism, not promising a health outcome or a fixed staffing reduction.
The metal-textile study reports an initial workload reduction above 50 percent
S34
The shift envelope still needs explicit limits. The inherited Cameroon labour entry records a standard 40-hour week and a 48-hour maximum.[R16] Vision does not make sustained exception review harmless within those limits. Work design should include rotation between verification, physical confirmation and line coaching; recovery pauses away from the display; and a cap on repeated alarms that triggers supervisor intervention. The relevant control is the pattern of demands across the whole shift, including the late period when earlier fatigue becomes visible.
The inherited Cameroon labour entry records a standard 40-hour week and a 48-hour maximum
R16
Exhibit 56 shows an illustrative allocation of a 400-minute net work window. It is a design case, not a time-and-motion study.
Exhibit 56 shows an illustrative allocation of a 400-minute net work window
illustrative
Read-out
Illustrative task allocation, not measured GFT performance. Both cases total 400 minutes after breaks and handover. Manual scanning is assumed to fall from 320 to zero; exception review is set at 160 minutes, a 50 percent reduction of the former scanning load. The released 160 minutes are assigned to coaching and to optics and records. Actual allocation requires floor measurement.
Unit: minutes per 400-minute net work window. Status: illustrative. Sources: S34, S07. Illustrative figures are teaching examples, not measured factory results.
In this illustrative allocation, the arithmetic is intentionally simple. The triage case reserves 160 minutes for exceptions, half of the former 320-minute scanning load, then assigns the same amount to coaching and to optics and records. This preserves total paid time and demonstrates why the proposal is a task redesign rather than an automatic headcount case. The actual queue may be larger or smaller. If exceptions consume the whole window, the correct response is to investigate forwarding and process conditions, not to accelerate clicking.
Physical layout determines whether the new task is sustainable. The monitor should sit at a comfortable viewing height and distance without forcing neck rotation away from the material. Controls should allow seated and standing work where the station permits. Ambient light should not wash out the anomaly display. The inspector needs direct access to the physical roll or garment, because some decisions depend on handling, transmitted light or comparison with a reference sample. Touch targets must remain usable with the gloves and hand positions required at that station.
Alarm design deserves the same engineering care as camera placement. Repeated audible alerts should be reserved for conditions that require immediate containment. Routine exceptions can enter a visible queue. Alerts should combine into one event when they refer to the same physical defect, and the interface should distinguish an image-quality failure from a suspected product defect. A supervisor needs to see queue age and station health, not just the number of red boxes.
Finally, the factory should test whether people are actually better off. Short confidential pulse checks, observation of posture, late-shift review quality and worker suggestions can reveal burdens that system logs miss. Retention and job satisfaction may be affected by many causes, so the deployment should not attribute changes to vision without a suitable evaluation. The defensible initial claim is narrower: triage can remove continuous searching and make workload visible, provided the factory acts on what it measures.
Interlude · the cloth half-woven
Woven: sections 12.0 to 12.3Still to weave: 12.4 to 12.7
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 ↗12.4
A thin supervisor pipeline should be treated as a greenfield staffing assumption to be tested during recruitment, not as a quantified national fact. The inherited labour entry reports that about 90 percent of employment in Cameroon is informal, but that statistic does not count garment supervisors or establish their competence.[R16] It does show why a formal factory cannot assume that every applicant arrives with experience of structured quality routines, documented escalation or digital work instructions.
The inherited labour entry reports that about 90 percent of employment in Cameroon is informal
R16
Vision can reduce the amount of abstract explanation a new supervisor must give. A skipped stitch, an open seam or a puckered seam can be shown as a cropped image beside an accepted example and the physical item. The anomaly map helps locate the region, while the defect dictionary supplies the class and severity. The conversation moves from a general accusation of poor workmanship to a specific question about thread tension, needle condition, fabric handling or method.
Good coaching follows a disciplined sequence. Confirm that the image and the physical item refer to the same unit. Ask the operator to describe what changed. Check the most plausible machine, material, method and environment causes. Agree one corrective action and its owner. Return to later evidence to see whether the action held. The camera supplies a shared object for discussion; it does not supply the root cause by itself.
This is a different supervisory job from watching a dashboard. The line quality lead must understand the process well enough to distinguish a recurring mechanism fault from an isolated handling error. The lead must also know when the image is misleading. A new print can trigger anomalies without any change in sewing. A tension problem can be real even when the model fails to flag it. Coaching therefore combines product knowledge, evidence interpretation and the authority to stop work.
Worker voice improves this learning loop. International Labour Organization case studies distinguish AI uses that complement skills from those that replace or deskill them, and identify worker oversight over how tools organise training and work as an important condition for greater discretion.[S1201] Supervisors should consequently review defect examples with operators before changing standards or incentives. An operator's explanation can reveal a material or method cause that no image contains.
The same principle protects dignity. A defect thumbnail is evidence about a product at a time and station, not a verdict on a person's effort or potential. Coaching should begin with process conditions and use individual attribution only where the evidence supports it. Repeat defects at one station call for examination of machine settings, work instructions, training coverage and incoming material before disciplinary action is considered.
A practical coaching board needs only a small set of views: the defect example, its class and severity, the agreed physical cause, the corrective action and the verification result. Keeping this view at line level prevents the quality lead from disappearing into central reporting. It also creates a teaching library in which new supervisors learn from resolved cases rather than relying solely on an experienced person's memory.
Supervisor assessment should reproduce that complete coaching cycle. An assessor selects a real or deliberately seeded case, asks the candidate to explain what the image can and cannot establish, and observes whether the operator is invited to contribute. The candidate must then choose a proportionate containment action and define the later evidence that would confirm recovery. A technically correct diagnosis is insufficient if the candidate humiliates the operator, hides uncertainty or closes the action without verification. This assessment makes respectful evidence handling a supervised skill rather than a statement of values.
12.5
Visual Inspector
Visual Inspector: Assigned decision autonomy 2, Required competence breadth 4 · derived
The proposed career structure gives each new responsibility a name, an authorisation boundary and an assessed competence.
Vision Station Operator
Vision Station Operator: Assigned decision autonomy 4, Required competence breadth 6 · derived
Quality Analytics Technician
Quality Analytics Technician: Assigned decision autonomy 6, Required competence breadth 8 · derived
Line Quality Lead
Line Quality Lead: Assigned decision autonomy 8, Required competence breadth 8 · derived
Vision Lead
Vision Lead: Assigned decision autonomy 10, Required competence breadth 10 · derived
Derived planning rubric, not measured attainment. Horizontal position = 2 x assigned autonomy level from 1 to 5. Vertical position = 2 x required competence-domain count from 1 to 5. Domains are product quality, station operation, data curation, coaching and model governance. Role assignments are GFT design assumptions informed by exception inspection, formal work organisation and worker-discretion principles; they require local validation.
Unit: GFT ordinal role-design score. Status: derived. Sources: S34, S1201, R16.
12.5 · the bodyExhibit 57 above; its figures are illustrative where marked.
The proposed career structure gives each new responsibility a name, an authorisation boundary and an assessed competence. It is a role architecture for recruitment and training, not a statement that these positions already exist at GFT or at a client factory. Progression depends on demonstrated judgement, not on willingness to accept automated decisions.
Exhibit 57 translates that architecture into an ordinal planning rubric. It does not represent observed skill scores, pay grades or a promised establishment.
The rubric makes an important restriction visible. A person may be an excellent garment inspector and still need preparation before governing a model. Promotion into analytics or vision leadership therefore requires evidence in the additional domains, not just seniority. Conversely, a technician who understands confusion matrices but cannot judge a seam should not hold final product-release authority. The two career paths can meet without collapsing into one generic digital role.
Role redesign should protect employment relationships during learning. The wage evidence argues against presenting the change as a simple labour-substitution exercise: the inherited entry records Cameroon's minimum wage at XAF 43,969 a month from 23 February 2024.[R14] The wider paper locates the economic case in avoided waste and rework. The workforce case is that experienced inspectors carry defect knowledge which the system needs, while new technical responsibilities create a route for that knowledge to be developed rather than discarded.
the inherited entry records Cameroon’s minimum wage at XAF 43,969 a month from 23 February 2024
R14
Workers and their representatives should help define the transition. Consultation can test whether the interface removes useful discretion, whether training is accessible and whether exception data could become covert individual surveillance. The International Labour Organization research links better outcomes to worker voice, skill investment and transparent routes to challenge technology-enabled decisions.[S1201] A formal appeal route is therefore part of the operating design: an inspector can request review of a disposition, and an operator can challenge an inference made about their work.
Recruitment should assess practical reasoning before software familiarity. Can the applicant compare a physical defect with a standard, explain uncertainty and stop a process when evidence conflicts? Those behaviours are harder to improvise after launch than basic touchscreen operation. Existing inspectors should receive the first opportunity to enter the new roles, with bridging instruction rather than an assumption that digital recruits are inherently superior.
12.6
The bootcamp should be a supervised floor programme, not a slide course followed by a login. An illustrative full-time format uses the standard 40-hour week recorded for Cameroon, giving 4 x 40 = 160 learning hours across four weeks.[R16] This duration is a planning assumption, not an externally validated certification standard. Shift cover, prior attainment and the availability of a competent instructor may justify a longer pathway.
An illustrative full-time format uses the standard 40-hour week recorded for Cameroon
R16 · illustrative
Exhibit 58 defines the proposed modules and observable competence gates.
Read-out
| Week 1: Product quality and triage | Grade accepted and defective samples; route critical, uncertain and nominal cases | Correct containment and escalation on the assessor's complete case set |
|---|---|---|
| Week 2: Optics and station care | Clean optics; check reference tile, illumination, encoder and fallback | Checklist completed; failed reference check correctly stops automatic PASS |
| Week 3: Evidence and active learning | Disposition crops; resolve disagreements; read a confusion matrix | Every sampled crop carries valid class, severity, context and uncertainty |
| Week 4: Coaching and control | Lead a defect review; test a cause; run rollback and appeal drills | Quality lead accepts observed coaching, containment and review decisions |
Illustrative GFT curriculum based on a 40-hour standard week and cited inspection and worker-voice principles. It is not an existing GFT course, an accredited qualification or evidence of staff attainment. A local quality authority must approve content, language, assessment and any use of the term certification.
Unit: illustrative full-time learning programme. Status: illustrative. Sources: R16, R29, S34, S1201. Illustrative figures are teaching examples, not measured factory results.
The first module establishes the product standard before the model. Trainees learn the paper's defect dictionary, severity rules and the difference between a defect, an anomaly and an uncertain image. The inherited evidence describes computer vision across more than 40 defect types, but that breadth does not remove the need to define local acceptance criteria for each product.[R29] Assessment uses physical samples and deliberately includes borderline cases.
The inherited evidence describes computer vision across more than 40 defect types
R29
The second module makes station care a quality responsibility. Trainees practise lint removal, reference-tile checks, illumination inspection, encoder verification and controlled fallback. They learn to distinguish a dirty viewport from a process defect. They also learn what they may not do: cleaning must not become informal recalibration, and a supervisor must not silence a failed check to keep production moving.
The third module develops evidence judgement. Trainees review false alarms and escapes, complete context fields, separate duplicate defects and interpret a confusion matrix by class. Group review exposes disagreement and teaches when to seek consensus. The assessment rewards correct escalation of uncertainty, not speed or agreement with the model.
The fourth module moves the trainee into supervision. Each candidate leads a case review, asks for operator evidence, proposes a cause, agrees a corrective action and checks a later result. Rollback, manual fallback and appeal drills confirm that authority is real. A quality lead observes performance and records remaining development needs. Passing the course means demonstrated competence on the defined cases, not permanent mastery of every fabric and model.
Training continues after the bootcamp through short case reviews, cross-training and refreshers triggered by new products or material model changes. Each station needs more than one person able to perform safe fallback. Supervisors also need separate practice in interpreting queue and drift measures, because an operator's competence does not automatically prepare a lead to govern the whole decision cell.
12.7
Related · chapter 019 paragraphs step it
Read-out
Derived ordinal coding, not source measurements. Dependency: 0 none, 1 current nominal images, 2 several records, 3 multi-period history. Readiness: 0 exploratory, 1 weakly validated, 2 assisted decision, 3 inspectable physical decision. Vision still needs nominal local frames. Forecasting and inventory need accumulated transactions. Generative design needs specialist product validation.
Unit: GFT analytical score, 0 low to 3 high. Status: derived. Sources: R27, R29, R30, R31, S04, S05, S31.
Read this exhibit in its chapter ↗People measures should show whether the decision cell is becoming more capable, not whether workers are becoming more compliant. The Performance-Management System should connect station evidence to skill-matrix coverage, current training authorisation, review quality, coaching closure and safe escalation. Each measure needs a named denominator and a clear owner. A percentage without those definitions can make a weak training system appear strong.
Skill-matrix coverage counts people whose observed competence is current for a defined role and product family. It should distinguish course attendance from successful assessment and should show backup coverage for fallback. Recheck agreement compares a sample of dispositions with an independent quality review, including uncertainty and escalation. Coaching closure tracks whether agreed actions have an owner and a later verification, rather than counting meetings held.
Exception burden belongs beside those capability measures. Queue age, repeated alarms for one physical flaw and after-hours spillover reveal whether the system is creating hidden work. Appeal outcomes show whether challenges change a disposition or expose a defect in the standard. Bypass and fallback records show whether people can operate safely when the model or optics fail. These measures make resistance diagnostically useful instead of treating it as a behavioural problem.
A small measurement dictionary should resolve recurring disputes before the dashboard is built. Recheck agreement needs a sampling rule that includes accepted and rejected items, not only cases selected by the model. Coaching closure needs a due condition and evidence of effect, not a checkbox marked after discussion. Training currency needs an expiry rule linked to product or model change. Exception burden needs to separate product alarms from station-health alarms. Without these definitions, different shifts can improve their reported position simply by counting differently.
The quality authority should own recheck sampling, the people lead should own competence records, and the production lead should own removal of queue and workload constraints. Shared review matters because no single function can interpret the measures alone. A rise in disagreement may reflect weak training, a changed product standard or model drift. A fall in exceptions may reflect better screening, a failed camera or inspectors bypassing the queue. The KPI review must test those competing explanations against floor evidence.
The KPI design must also restrict misuse. Raw defect counts at an operator's station are not a fair individual score because machine condition, material, style difficulty and sampling influence them. Model agreement is not a fair inspector score because the model can be wrong. Individual images and event logs should not be repurposed for discipline without a defined rule, access control and an opportunity to contest. This follows the International Labour Organization's warning that algorithmic management can intensify monitoring and reduce worker control when oversight is weak.[S1201]
The management cadence should close with capability action. The daily huddle resolves exceptions and coaching needs. The weekly quality review examines recheck disagreement, queue burden and repeat causes. The monthly people review updates training authorisation, backup coverage and role progression. Where evidence shows overload or workarounds, management changes the system rather than asking inspectors to compensate indefinitely.
This is how vision elevates supervision. Inspectors become authors of verified evidence. Technicians become stewards of data quality. Supervisors become coaches with time and visual proof to improve the process. The model remains subordinate to a social and technical system that can question it, maintain it and stop it.
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 · 24 citation marks on this pageMost cited: S34, 5 times
Grade AGrade BGrade CKnot size: citations on this page
All 7 records shown
R14 · Grade C · inherited record
Inherited evidence R14
Cameroon minimum wage (SMIG) XAF 41,875 a month in 2023, raised to XAF 43,969 a month from 23 February 2024, about US$76, a rise of 5 percent.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R16 · Grade C · inherited record
Inherited evidence R16
Informal share of employment about 90 percent; standard 40-hour week, 48-hour maximum, overtime at 1.2 times.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R29 · Grade C · inherited record
Inherited evidence R29
Manual defect inspection 60 to 70 percent accurate, missing 20 to 30 percent of defects; computer vision above 90 percent across more than 40 defect types; line-scan inspection up to 60 metres a minute.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
c12-S03 · Grade B
Fabric Defect Detection Using Computer Vision Techniques: A Comprehensive Review
Lack of concentration, human fatigue, and time consumption are the main drawbacks associated with the manual fabric defect detection process.
2020-11-16. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S07 · Grade B
Low Cost Two-Stage Fabric Defect Detection at the Edge
We further show that forwarding here is false-positive-limited rather than prevalence-limited - 85% of forwarded frames are false alarms - and quantify the 29-45% inference reduction attainable under tighter calibration.
2026-08-12. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S34 · Grade B
Anomaly Detection in Metal-Textile Industries
Only cases of doubt where an anomaly has been detected are inspected by the human inspector.
2025-01-03. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S1201 · Grade A
Global case studies of social dialogue on AI and algorithmic management
Alternatively, these tools can complement workers' existing skills or help them to develop new skills and modes of working - augmenting rather than replacing workers.
2025-07. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗S34 IntechOpen. 2025-01-03.
S1201 International Labour Organization. 2025-07.
R29 MDPI; industry inspection literature. Publication date unrecorded; inherited source-text check.
S07 arXiv. 2026-08-12.
S03 Mathematical Problems in Engineering. 2020-11-16.
R16 WageIndicator. Publication date unrecorded; inherited source-text check.
R14 WageIndicator. Publication date unrecorded; inherited source-text check.
Keep the thread · chapter 12 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.