Research librarySeeing Every Defect
Chapter 11 of 14 · Part III · Draft web edition
Counting defects without fixing their root cause merely automates scrap accounting; closed-loop integration wires edge cameras directly into the Performance-Management System.
The cloth is woven from this chapter's 5 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 system design derived from cited vision triage, digital quality, DMAIC and SQC methods. It is a proposed architecture, not evidence of an implemented GFT result.
Unit: proposed operating flow. Status: derived. Sources: R29, S07, S24, S25, S26.
Overture · 1 min
Computer vision improves a garment factory only when a detection changes a decision before the same fault repeats. The destination is therefore not a model dashboard. It is the Performance-Management System, where a verified camera event becomes a defect count, an hourly exception, an assigned action and, eventually, evidence that the action worked. GFT's recommended closed loop has four controls: one quality-event schema from manual entry and camera inference, statistical signals that distinguish a process shift from ordinary variation, management cadences with named decision rights, and upstream feedback to cutting, maintenance and suppliers. A bounding box without those controls is only a more precise record of loss.
This chapter treats integration as an operating design rather than a software connection. It defines what the edge node should write, when a person must confirm it, how Defects Per Hundred Units and Right First Time remain comparable, and which meeting may change which setting. It also separates process control from customer acceptance. A stable process can still make unacceptable garments, while a noisy process can occasionally pass a final audit. The Performance-Management System must show both conditions rather than hiding them behind one score.
11.1
Read-out · sourced
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.
The programme's six KPI families, Production, Quality, Delivery, Cost, People and Sustainability, form one management record. Vision does not need a seventh family or a parallel scorecard. A verified sewing defect belongs in Quality. Its rework minutes affect Cost and People. A late corrective action affects Delivery. Fabric avoided through marker placement affects Sustainability. The same event can inform several families, but it should be captured once and referenced by a common identifier. That discipline prevents the camera supplier, quality lead and production manager from reporting different versions of the same shift.
The destination also determines the latency required. A model confidence score may be useful to a vision technician, but it cannot lead a line huddle. The huddle needs a small set of decisions: which defect rose, at which station, on which style or fabric lot, who is investigating, and whether production may continue. The event contract must therefore preserve operational context alongside the image result. Camera coordinates without a bundle identifier cannot guide cutting. A defect class without a station identifier cannot guide maintenance. A thumbnail without the final human disposition cannot be treated as confirmed truth.
Published garment evidence supports this management integration, while also warning against a simple technology claim. In a Central Java garment study, digital quality control had a reported path coefficient of 0.262 with p = 0.002, production automation had a coefficient of 0.639 with reported p = 0.000, and the two predictors jointly explained 64.5 percent of product-quality variance. These are study-specific associations from a questionnaire and structural model, not guaranteed effects for a Cameroonian factory. They support combined measurement and production control, not unattended camera acceptance.[S24]
Exhibit 50 keeps those empirical anchors beside two other garment studies. Their measures are not interchangeable. Path coefficients, defect shares and defective-unit counts answer different questions. The exhibit deliberately retains those definitions instead of converting them into a single performance score.
A separate Six Sigma garment study found that broken threads and puckering together represented 23 percent of 16 defect types. The value of that finding is not the percentage itself. It is the management sequence: Pareto analysis concentrated attention, fishbone analysis organised possible causes, and the improve stage assigned countermeasures. Vision can make the Pareto current and visual, but the root-cause work still determines whether recurrence falls.[S25]
11.2
Exhibit 5211 paragraphs step it
Read-out
| Units inspected | units_inspected | Quality | Supervisor checks denominator hourly |
|---|---|---|---|
| Defect count and class | defect_count, class_code | Quality | Quality lead reviews Pareto daily |
| Units without rework | rft_numerator | Production, Quality | Line lead reviews RFT hourly |
| Repair minutes | rework_minutes | Cost, People | Production manager reviews daily |
| Late action flag | action_due_state | Delivery | Plant review escalates weekly |
| Fabric avoided or scrapped | fabric_disposition_mass | Sustainability, Cost | Cutting and purchasing review weekly |
Illustrative GFT schema and decision-rights design. S24 supports integrated digital quality management generally, but does not specify these fields or cadences.
Unit: illustrative PMS field mapping. Status: illustrative. Sources: S24. Illustrative figures are teaching examples, not measured factory results.
The manual crawl phase and automated phase must write the same logical record. During manual operation, a quality lead enters units inspected, defective units, defect counts, defect class, station, style and disposition. When vision starts, the edge node supplies the detection fields and a person verifies exceptions according to risk. The storage technology may change from workbook to local buffer and central database, but the definitions of numerator, denominator, time period and responsibility must not change. Otherwise the apparent improvement could be nothing more than a change in accounting.
Three measures need explicit separation. Defects Per Hundred Units is total verified defects divided by units inspected, then expressed on a per-hundred-unit basis. It allows several defects on one garment. The defective-unit proportion is garments with at least one verified defect divided by garments inspected. Right First Time is units completed without repair or rework divided by units completed, expressed as a percentage. A garment with two defects counts twice in DHU but once in the defective-unit proportion. Mixing these measures would corrupt both the p-chart and the incentive calculation.
The recommended quality event carries a timestamp, roll or bundle identifier, garment or style identifier where available, station, defect class, severity, model confidence, image reference, coordinates, verifier disposition and final physical disposition. Model confidence is metadata, not a quality KPI. The verifier disposition records whether the detection was accepted, rejected or left pending. The physical disposition records what happened to the item: pass, repair, recut, downgrade or scrap. Keeping those two decisions separate stops a correctly detected fault from being mistaken for a completed corrective action.
An edge study of two-stage fabric inspection explicitly positions its system as assisted triage rather than autonomous acceptance. That boundary matters for the Quality Log. Clearly acceptable frames may pass under an approved rule, while flagged or ambiguous cases receive human review. The log must then record the review state, because unreviewed model alarms are not equivalent to verified defects.[S07]
The inherited inspection register summarises manual defect detection at 60 to 70 percent accuracy and computer vision above 90 percent across more than 40 defect types. Even where a factory reproduces that detection gain, it describes classification performance rather than process improvement. The PMS must show whether verified detections lead to containment, corrected settings and fewer repeat events. Otherwise a better detector can simply produce a longer list of unresolved faults.[R29]
Exhibit 51 shows the proposed path from image to controlled process. The return path is essential. If an action does not restore the signal, the event remains open and escalates. If it does restore the signal, the closure evidence becomes part of the control plan.
Exhibit 52 translates that flow into a compact field contract. It is an illustrative GFT design, not a claim that a factory has already implemented these fields. Its purpose is to make ownership and review cadence inspectable before software procurement begins.
The schema also needs rejection rules. An event with an unknown station, an impossible timestamp, a missing denominator or an unresolved bundle should enter an exception queue rather than silently distorting the KPI. Late-arriving buffered events should retain their original production timestamp and carry a separate ingestion timestamp. That distinction lets the hourly chart reconstruct the actual production period while the interface honestly shows that synchronisation occurred later.
A quality event needs a lifecycle that everyone can read. A model-generated detection begins as unverified. Human disposition can confirm it as a defect, reject it as a false alarm or hold it for further inspection. A confirmed defect then receives a physical disposition and, where action is required, a linked corrective-action record. Closure is a later state supported by process evidence. Collapsing these states into a single detected flag would inflate confirmed DHU, hide pending work and make false alarms look like production failures.
State transitions also protect offline operation. Each edge event needs a durable unique identifier assigned at capture, so store-and-forward synchronisation cannot insert it twice. If a verifier changes a disposition, the system should append a revision with author, time and reason rather than overwrite the original decision. The KPI service reads the latest valid revision, while auditors can reconstruct why the measure changed. This is especially important when several stations reconnect after a network interruption and historical hourly buckets must be recalculated.
The Quality Log should publish its completeness state beside its values. An hourly chart based on verified events alone may look reassuring while a large verification queue remains unseen. The supervisor therefore needs counts for captured, verified, rejected and pending events, plus any periods with missing denominators. Management can then distinguish a genuine low-defect hour from an hour whose evidence is incomplete. This small governance feature prevents operational pressure from turning missing review into apparent quality.
11.4
H01
H01: Defective units 2, Baseline centre line 2.5, Upper control limit 5.8 (percent of units defective by hour) · illustrative
A live defect total is not yet a control signal.
H03
H03: Defective units 2.1, Baseline centre line 2.5, Upper control limit 5.8 (percent of units defective by hour) · illustrative
H05
H05: Defective units 2.3, Baseline centre line 2.5, Upper control limit 5.8 (percent of units defective by hour) · illustrative
H07
H07: Defective units 2.2, Baseline centre line 2.5, Upper control limit 5.8 (percent of units defective by hour) · illustrative
H08
H08: Defective units 2.5, Baseline centre line 2.5, Upper control limit 5.8 (percent of units defective by hour) · illustrative
H10
H10: Defective units 2.9, Baseline centre line 2.5, Upper control limit 5.8 (percent of units defective by hour) · illustrative
H12
H12: Defective units 3.1, Baseline centre line 2.5, Upper control limit 5.8 (percent of units defective by hour) · illustrative
The method behind it: Worked example: one hourly andon and three distinct KPIs.
Illustrative model. Each hour has n = 200 inspected units. The assumed stable baseline is p-bar = 0.025. UCL = 0.025 + 3 x square_root(0.025 x 0.975 / 200) = 0.0581, shown as 5.81 percent. LCL is floored at zero. Hour 11 is 6.5 percent and signals; the other constructed points remain below the limit.
Unit: percent of units defective by hour. Status: illustrative. Sources: S1101. Illustrative figures are teaching examples, not measured factory results.
11.4 · the bodyExhibit 53 above; its figures are illustrative where marked.
A live defect total is not yet a control signal. Production volume changes by hour, and several defect classes can rise together when a new style enters the line. The p-chart addresses one precise question: is the proportion of defective units reasonably consistent with the established process, or has something changed? NIST defines the proportion nonconforming as nonconforming items divided by all items inspected and grounds the p-chart in the binomial distribution. The chart is therefore built on defective units, not total defects.[S1101]
For subgroup i, the system calculates p_i = D_i / n_i, where D_i is verified defective units and n_i is units inspected. With a stable baseline proportion p-bar, the upper control limit is p-bar plus three times the square root of p-bar(1 - p-bar) / n_i. The lower limit uses the same term with subtraction and is floored at zero. When hourly volume changes, the limits change with n_i. This matters on a garment floor: fifty defective units has a different meaning in a small hour than in a large one, and the chart must retain that denominator.[S1101]
Control limits are not specification limits. A point inside the limits suggests a stable process; it does not prove that the defective proportion satisfies a buyer. Conversely, a point outside the limits calls for investigation even if the final sampled audit happens to pass. The andon reaction should therefore be triggered by process evidence, not postponed until an end-of-line acceptance decision. This separation also prevents a perverse response in which supervisors widen control limits merely to reduce alarms.
The baseline itself needs governance. It should come from a defined reference period with comparable products, verified dispositions and operating conditions. Known special causes, such as a failed needle or an uncalibrated light, should be investigated before their data are used to describe common-cause variation. After a confirmed process improvement, the team may recalculate the centre line under change control. It should never redraw limits continuously around a deteriorating process, because that would normalise the deterioration.
Exhibit 53 illustrates the mechanism with a fixed hourly denominator and a baseline from an earlier stable period. Only the eleventh hour crosses the upper limit. The correct response is not to declare the whole shift defective. It is to contain suspect work, inspect the station and material context, and record whether the next verified observations return to control.
A useful andon policy also suppresses noise without hiding risk. One high-severity critical defect may require immediate containment regardless of statistical significance. Repeated minor defects may require a run-rule or a confirmed class-specific pattern before stopping the line. A high false-alarm rate may indicate an optical or model problem rather than a production shift. The response matrix should therefore combine severity, statistical evidence and verification state. It should not reduce every condition to the same red light.
The Indonesian SQC and FMEA case provides a relevant warning. Its p-charts still showed points outside the control limits for faded and misaligned prints, leading the authors to judge that the process was not fully under statistical control. The study then used risk priorities and work-standard recommendations. The operational lesson is the sequence from signal to cause to action, not the transfer of its defect rates to another factory.[S26]
Interlude · the cloth half-woven
Woven: sections 11.0 to 11.4Still to weave: 11.5 to 11.8
Read-out
Sourced reference values from different studies and datasets. R29 is the inherited inspection summary. S07 concerns a four-class knit-fabric edge cascade. S17 concerns a sewing-model ablation. S19 reports an Ethiopian garment study and an estimated return, not an achieved GFT payback.
Unit: reported research and inherited reference values. Status: sourced. Sources: R29, S07, S17, S19.
Read this exhibit in its chapter ↗11.5
Related · chapter 127 paragraphs step it
Read-out
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.
Read this exhibit in its chapter ↗Closed-loop control fails when every exception is sent to every meeting. Hourly, daily, weekly and monthly reviews have different decision rights. The hourly line conversation protects current production. The daily huddle removes recurring obstacles and confirms ownership. The weekly plant review tests causes across lines and authorises changes to maintenance, work methods or supplier handling. The monthly sales and operations planning review addresses capacity, purchasing and product-mix consequences. Escalation should move an unresolved decision upward, not merely repeat the same chart in a larger room.
The hourly review is deliberately narrow. The supervisor sees the current andon, the affected station, the verified thumbnail and the suspect bundle. The team chooses containment, immediate repair or continued observation under a defined rule. It records the action and expected verification time in the same event thread. If no one can act, the exception escalates with its evidence intact. The camera has then shortened the path from detection to responsibility without pretending to diagnose the cause itself.
The daily line huddle, planned as a 15-minute illustrative operating routine, reviews the prior period's largest verified classes, unresolved actions and denominator quality. Thumbnails help the team discuss the same physical condition, but the chair must ask what changed in the process. A rise in skipped stitches may prompt a needle and tension check. A rise in shade exceptions may prompt a roll-lot and illumination check. The huddle closes only actions with verification evidence, not actions with a completed meeting note.
The weekly review compares lines, styles, shifts and fabric lots under consistent definitions. It is the natural home for DMAIC selection and FMEA revision because those tasks require cross-functional evidence. A mechanic, quality lead, production supervisor and cutting representative can test whether the same defect follows a machine, material lot or method. The monthly review then uses confirmed trends for supplier discussions and capacity planning. It should not chase individual hourly points that the line has already resolved.
A simple action record prevents meeting theatre. It contains the originating quality-event identifiers, suspected cause, chosen countermeasure, accountable role, due time, containment status and verification result. Reopening is preferable to a falsely closed action. If the same class returns after closure, the system links the recurrence to the earlier action so the weekly review can reject the presumed cause rather than repeat it.
Decision rights should be written before the first alarm. Operators may contain and perform approved immediate corrections. Supervisors may authorise a line-specific response within the reaction plan. Maintenance changes to controlled settings require the designated technical role. Supplier claims, specification changes and production-plan trade-offs belong with the appropriate plant or commercial review. This allocation prevents a huddle from making decisions it cannot sustain and prevents urgent exceptions from waiting for a meeting with no operational authority.
Each cadence should also publish a small unresolved queue, ordered by risk and age rather than by the latest timestamp. The chair begins with actions whose verification failed or whose containment remains open. New Pareto entries come afterwards. That ordering stops a fresh stream of images from displacing yesterday's unfinished corrective work. It also creates an honest escalation route when the same constraint crosses several management levels.
11.6
Units inspected to Direct first pass
Units inspected to Direct first pass: 92,000 illustrative garment-equivalent units · illustrative
Incoming fabric inspection creates value before sewing only when its coordinates survive the journey to the cutting table.
Units inspected to Sewing flags
Units inspected to Sewing flags: 4,000 illustrative garment-equivalent units · illustrative
Direct first pass to Accepted for packing
Direct first pass to Accepted for packing: 92,000 illustrative garment-equivalent units · illustrative
Fabric flags to Fabric downgrade
Fabric flags to Fabric downgrade: 1,000 illustrative garment-equivalent units · illustrative
Sewing flags to Sewing downgrade
Sewing flags to Sewing downgrade: 300 illustrative garment-equivalent units · illustrative
Sewing flags to Recut or scrap
Sewing flags to Recut or scrap: 200 illustrative garment-equivalent units · illustrative
Instant repair to Accepted for packing
Instant repair to Accepted for packing: 3,500 illustrative garment-equivalent units · illustrative
Illustrative conserved routing model using 100,000 garment-equivalent units to explain destinations, not an observed factory result. Input splits into 92,000 direct first-pass, 4,000 incoming-fabric flags and 4,000 inline-sewing flags. Marker avoidance returns 3,000 to packing, while 1,000 fabric units are downgraded. Instant repair returns 3,500 to packing, with 300 downgraded and 200 scrapped. Packing receives 98,500 units in total.
Unit: illustrative garment-equivalent units. Status: illustrative. Sources: R29, S25. Illustrative figures are teaching examples, not measured factory results.
11.6 · the bodyExhibit 54 above; its figures are illustrative where marked.
Incoming fabric inspection creates value before sewing only when its coordinates survive the journey to the cutting table. A line-scan detection should be stored against roll identifier and web position, with cross-web and along-web coordinates under a declared measurement convention. The cutting system then translates that roll map into the current spread and marker. An operator can splice a severe fault, shift a marker where the rules permit, or place an acceptable minor flaw outside a critical panel zone. The decision is physical and immediate, not a supplier report read after the bundle has been sewn.
The interface must respect changing material states. A roll becomes a spread, a spread becomes plies, and cut panels become bundles. Each transformation needs a new identifier linked to its parent, because roll coordinates alone no longer locate a panel after cutting. Splice removal, end loss and marker movement alter the mapping. The practical design is therefore a chain of transformations with explicit losses, not an image viewer attached to a roll. If the chain is broken, the system should mark the affected panels for verification rather than invent coordinates.
Severity rules should reflect product use. A small flaw that is acceptable in a hidden internal panel may be unacceptable on a visible uniform chest. The marker rule engine can encode those zones, but a cutting lead should approve exceptions during early deployment. This avoids two failure modes: discarding usable fabric through excessive caution and hiding known defects inside garments through aggressive nesting. Both outcomes would damage the economic case even if the detector remained accurate.
Upstream feedback also needs fair evidence. A supplier scorecard should separate confirmed fabric-origin faults from faults created during spreading, cutting or sewing. It should show inspected length, confirmed fault classes, severity and disposition before assigning a claim. Camera imagery strengthens the conversation, but it does not by itself prove when damage occurred. Joint review of roll condition, handling records and mapped coordinates is necessary where responsibility is disputed.
Exhibit 54 shows an illustrative routing balance. It is not a forecast of savings. Its purpose is to force every inspected unit-equivalent into a declared destination and to show that detection, avoidance, repair and packing are different events. A real deployment would replace the constructed quantities with reconciled PMS records.
The same conservation discipline belongs in the daily report. Units inspected should reconcile to first-pass acceptance, verified exceptions and unresolved items. Exceptions should reconcile to repair, downgrade, recut, scrap and still-open dispositions. If those quantities do not balance, management has a data-integrity problem before it has a quality-improvement result. The system should surface the imbalance rather than force a tidy total.
11.7
The weekly Pareto is a selection tool, not a root cause. It tells the team where observed quality loss is concentrated under the current classification rules. Define then states the problem with a product, station, period and operational defect definition. Measure verifies the denominator, image review and measurement repeatability. Analyse tests possible causes. Improve changes a condition under controlled observation. Control holds the new setting through standard work, ownership and a reaction plan. Skipping from a camera Pareto directly to a machine adjustment would automate guessing.
The jacket DMAIC study illustrates why this discipline matters. Its Pareto stage identified broken threads and puckering as the dominant categories, and its cause analysis considered human, machine, material, method and environment factors before recommendations were made. Vision can improve the speed and consistency of the Measure phase. It cannot decide which of those five factor families explains a particular recurrence.[S25]
A practical root-cause board should therefore place evidence beside hypotheses. If skipped stitches cluster after needle changes, the team examines needle specification, installation and thread compatibility. If puckering follows one fabric lot, it examines feed, tension, shrinkage and material behaviour. If false alarms cluster near one luminaire or after lens cleaning, it examines the measurement system before touching sewing settings. Each hypothesis needs a predicted observable effect and a result that would disconfirm it.
FMEA adds priority where the Pareto alone is insufficient. Frequency matters, but so do severity and detectability. A rare critical seam failure may deserve action before a common cosmetic fault. The SQC and FMEA garment study used Risk Priority Numbers to rank faded print, inappropriate colour and misaligned print, then linked recommendations to work standards, equipment inspection and operator accuracy. The transferable lesson is to connect statistical evidence and risk ranking, not to copy its scores.[S26]
RPN should not become false precision. Teams should retain the individual severity, occurrence and detection assessments, the scoring scale and the rationale. A small change in one subjective score can reorder a ranking. Where the evidence is weak, the review should commission a check rather than hide uncertainty in multiplication. Vision may improve detection scoring for the classes it reliably sees, but it says little about defects outside its field of view.
Control is the stage that proves integration. The team states which setting changed, which chart should respond, how long the observation window runs and who may approve a return to normal operation. A short-lived improvement followed by recurrence is an open experiment, not a closed DMAIC project. The PMS should preserve both outcomes so later shifts do not repeat failed remedies.
11.8
Overall Equipment Effectiveness multiplies Availability, Performance and Quality. Vision primarily strengthens the information behind Quality through consistent RFT and verified defect records, but its interventions can also affect the other terms. A line stop lowers Availability while it occurs. Exception handling may reduce running speed. Those effects should remain visible. A programme that reports only quality improvement while hiding stopped time would overstate the production benefit.
The causal accounting needs care. An andon stop may reduce this hour's Availability while preventing a larger rework queue later. The PMS should therefore retain the stop event, its reason and the subsequent verified outcome. Management can then compare immediate cost with avoided recurrence using actual production evidence. It should not assume that every camera stop was necessary or that every stop was harmful. The purpose is to improve the response policy over time.
Incentives create a second risk. Paying a line only for low recorded DHU can encourage reclassification, delayed verification or suppressed events. Paying only for output can encourage bypass of the andon. A fairer scorecard combines verified RFT, action closure, denominator completeness and agreed audit results. It distinguishes defects created on the line from incoming fabric faults and recognises prompt escalation of a genuine problem. The audit trail matters more than a single green dashboard tile.
Quality bonuses also need a measurement-system check. Supervisors should periodically compare camera dispositions, human verification and an independent sample under the same defect definitions. Disagreement is a reason to examine taxonomy, optics, training or model drift, not automatically to blame an operator. The next chapter addresses those human workflows in detail; here, the requirement is narrower: no incentive should depend on an unverified model score.
The closed-loop acceptance test is consequently behavioural and numerical. A planted or naturally occurring exception must reach the Quality Log with the correct denominator, create the intended andon, appear in the line review, generate an accountable action, and retain verification after intervention. The report must also reconcile units and expose unresolved items. Only then can leadership treat an improvement in DHU, RFT or OEE as an operating result rather than a change in what the software chose to display.
Evaluation should compare equivalent work, not merely an earlier week with a later one. Style complexity, fabric lot, order size and staffing can all change the opportunity for defects. The review should retain those covariates and use matched runs or a controlled pilot where practical. It should also report verification coverage and false-alarm burden beside RFT. Otherwise a favourable trend may reflect an easier mix or incomplete review rather than the corrective action. A claimed payback belongs only after those operating conditions and measurement checks have been made visible.
GFT's recommendation is to commission the loop before expanding model classes or camera count. One station with a reliable event contract, a working reaction plan and verified closure teaches the factory more than several disconnected detectors. Scale follows when supervisors use the record without duplicate entry, actions close with evidence, and upstream teams can trace their decisions back to the original detection.
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.
6 sources · 16 citation marks on this pageMost cited: S1101, 4 times
Grade AGrade BGrade CKnot size: citations on this page
All 6 records shown
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.
c11-S07 · Grade B
Low Cost Two-Stage Fabric Defect Detection at the Edge
We report this as a caution for cascade speedups measured without controlling the data path, and position the system as AI-assisted triage rather than autonomous acceptance.
2026-08-12. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S24 · Grade B
Digitalisasi Quality Control dan Otomatisasi Produksi: Analisis Pengaruhnya Terhadap Kualitas Produk Garment di PT X, Perusahaan Garment di Jawa Tengah
Both predictors jointly explain 64.5 percent of the variance in Product Quality, with automation showing the more dominant contribution.
2026-08-03. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S25 · Grade B
Quality control of garment product using DMAIC Six Sigma
The frequency of occurrence for these two types of defect reaches 23% of the total 16 types of defects.
2024-06-30. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S26 · Grade B
Analysis of Quality Control and Product Failure Risks in Garment Manufacturing Using Statistical Quality Control (SQC) and Failure Mode and Effect Analysis (FMEA) Case Study: CV Muharram Garment Indonesia
The results showed that out of 1,380 produced units, 107 units were defective, consisting of 43 faded-print defects, 36 misaligned-print defects, and 28 inappropriate-color defects.
2026-09-25. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S1101 · Grade A
NIST/SEMATECH e-Handbook of Statistical Methods: Proportions Control Charts
The underlying statistical principles for a control chart for proportion nonconforming are based on the binomial distribution.
2026-04-22. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗S24 CEMERLANG: Jurnal Manajemen dan Ekonomi Bisnis. 2026-08-03.
S25 OPSI. 2024-06-30.
S26 Jurnal Teknik Industri (JATRI). 2026-09-25.
S07 arXiv. 2026-08-12.
R29 MDPI; industry inspection literature. Publication date unrecorded; inherited source-text check.
S1101 National Institute of Standards and Technology. 2026-04-22.
Keep the thread · chapter 11 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.