Research librarySeeing Every Defect
Chapter 02 of 14 · Part I · Draft web edition
Understanding why human eyes miss defects on moving fabric explains why adding more manual inspectors cannot solve a factory's quality problem.
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
Illustrative scenario, not observed factory data. Continuous manual capture starts at 70 percent, within the inherited 60 to 70 percent range (R29), then declines over two hours. The rotation series resets after each 30-minute task change to test Sandia's recommended maximum continuous inspection block (S0205). Textile evidence also reports difficulty sustaining focus beyond ten minutes (S03). The 91 percent vision reference represents the inherited above-90-percent benchmark, not a measurement or guaranteed local result (R29).
Unit: illustrative percent of true defects captured. Status: illustrative. Sources: R29, S03, S0205. Illustrative figures are teaching examples, not measured factory results.
Manual inspection cannot provide a reliable quality floor merely by covering every roll or garment. Coverage describes how much product passes in front of an inspector. Capture describes how many of its defects that inspector actually finds. Textile literature reports manual accuracy of 60 to 75 percent, while the inherited programme register records a 60 to 70 percent range and defect losses of 20 to 30 percent under operating conditions.[S03][S08][R29] These are reported ranges from different sources, not a universal law and not exact complements that should be forced into one equation. They nevertheless support one operational conclusion: a manual gate can process all the cloth and still let a material share of its flaws reach cutting, sewing or packing.
The causes sit at four levels. Visual search becomes less reliable when a repetitive web moves quickly, defects are rare and attention has been sustained without a break. Lighting can improve visibility, but excessive brightness or glare can defeat the same task. Human decisions also vary around borderline defects, so two inspectors need not assign the same class or severity. Finally, four-point roll grading and Acceptable Quality Limit sampling convert imperfect observations into decisions with different blind spots.[S0205][S0204][S0202][S0203] Adding inspectors may increase throughput, but it does not remove those mechanisms.
For Green Future Tech Limited's Cameroon textile programme, prepared 31 January 2024 with sources refreshed in 2026, the correct response is therefore not to discard manual measurement. It is to measure its error before computer vision is introduced. The manual baseline must record speed, time on task, lighting, defect class, inspector agreement and downstream escapes. That baseline can then be compared honestly with an automated gate whose inherited benchmark is above 90 percent detection at web speeds up to 60 metres a minute.[R29] The comparison remains a design target to verify on local cloth, not a promised factory result.
2.1
Related · chapter 069 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 ↗Visual inspection is not one act of looking. The inspection model reviewed by Sandia National Laboratories separates search from decision. The inspector first searches for a possible signal, then compares it with a standard and responds. Most information during search is gathered during fixations, which average about 300 milliseconds in the reviewed literature.[S0205] On a moving fabric web, the next section of cloth arrives whether or not the inspector has completed that search and decision sequence. Pacing therefore removes time from both stages.
The repetitive structure of woven and knitted cloth makes the task harder, not easier. A regular background offers few natural places to stop, while the signal may be a small break in that pattern. The Sandia review reports that vigilance losses tend to be worse for rare events, difficult detection, repetitive tasks and situations without performance feedback.[S0205] Those conditions describe a long run of nominally acceptable fabric interrupted by an occasional fault. Training can improve search strategy and knowledge of the defect catalogue, but it cannot make the incoming signal frequent or stop the web.
Textile-specific evidence points in the same direction. A comprehensive fabric-inspection review identifies training burden, ignored small defects, human effort and difficulty maintaining focus for more than ten minutes as drawbacks of manual inspection.[S03] A knitted-fabric study separately describes eye fatigue from continuous attention to an illuminated inspection board and notes that inspection quality changes with working time, experience and personal viewpoint.[S0203] These findings do not justify a single fatigue curve for every mill. They do invalidate the assumption that an inspector's detection rate is constant across a shift.
The wider inspection literature provides a time scale and a warning against false precision. Sandia reports that the decline in vigilance performance is typically complete 20 to 35 minutes into a session, with at least half of the final loss occurring in the initial 15 minutes. It also records inspection studies in which detection deteriorated by up to 40 percent in 30 minutes, while stressing that performance depends on the task and its conditions.[S0205] Those figures come from the broader inspection literature, not a calibration of Cameroonian cotton lines. They support short inspection blocks and local measurement, not a universal claim that every fabric inspector fails at the twentieth minute.
A practical rotation follows from that evidence. Sandia's principal recommendation is to limit continuous inspection to 30 minutes, with a rest break or a change of task.[S0205] A factory can test that pattern without delaying production by rotating inspectors between roll watching, defect recording and exception review. The test should compare capture by time block on the same fabric family and defect mix. A simpler rota is useful only if the measured capture rate improves.
Lighting has a similarly conditional effect. The Sandia review identifies brightness, glare, colour and focus as the main lighting design variables. Increasing illumination can improve vision only up to a point, after which added illumination can create glare.[S0205] Manual fabric inspection commonly requires the operator to scan cloth transported across a brightly illuminated plane.[S0204] More backlight is therefore not automatically better inspection. The engineering question is whether each target defect retains sufficient contrast against its local weave, colour and surface relief.
Speed creates the second constraint. If an inspector has less time per unit area, the probability that a defect was never fixated rises. Sandia specifically identifies overly fast pacing as a cause of omissions because there is insufficient time for thorough inspection.[S0205] This is why a higher nominal roll throughput can produce a lower effective inspection coverage even when every metre passes the perch. The relevant measure is not metres presented per minute. It is the share of target defects found per metre, by class and contrast.
Exhibit 5 makes the distinction visible without pretending to offer measured factory curves. The manual values are an illustrative scenario anchored to the reported textile capture range. The break rotation is a test design, not a guaranteed recovery pattern. The vision line marks the inherited lower-bound benchmark above 90 percent and does not represent a locally validated result.[R29][S0205]
The exhibit cannot establish the slope at a particular factory. It shows why the baseline protocol must record time on task rather than pooling the whole shift into one percentage. It also shows why automation should be judged by sustained capture under the same optical and production conditions, not by a short demonstration on selected defects.
2.4
Finding a flaw and grading it are separate opportunities for variation. Signal detection theory, as summarised in the Sandia review, distinguishes perceptual sensitivity from response criterion. Sensitivity concerns the ability to separate signal from background noise. Response criterion concerns how much evidence an inspector requires before rejecting an item.[S0205] Two people can therefore see the same mark and make different decisions without either being careless.
That distinction matters on a production floor under conflicting pressures. A conservative inspector who rejects only clear defects will record fewer hits and fewer false alarms. A more lenient inspector will catch more borderline defects but also flag more acceptable cloth. The reviewed model predicts this movement between misses and false alarms as the decision threshold changes.[S0205] A single accuracy percentage hides which side of that trade-off a team has chosen.
Textile research explicitly describes manual inspection as subjective and inconsistent in high-speed manufacturing.[S01] Seam puckering provides a particularly clear example. The established grading method described in the research literature remains visual and subjective, while an automated alternative locates the seamline and extracts profile features before classifying the grade.[S21] The issue is not simply whether a wrinkle exists. It is where acceptable surface variation ends and a severity grade begins.
Similar ambiguity arises when a slub is close to a four-point length bracket or a stain is close to the background shade. A measurement taken after stopping and isolating the defect may differ from a judgement made on a moving web. The four-point method itself assigns points by fault length, so a boundary error changes the roll score even when the inspector detected the flaw.[S0203] Detection recall and grading agreement must therefore be measured separately.
A usable manual log needs three reference devices: a defect dictionary with photographs, boundary samples for each severity class, and a measured rule for length or position. The samples should include accepted as well as rejected cases. Otherwise, inspectors learn the centre of each category but not its edges. Periodic paired review then measures agreement on the same physical items, rather than comparing different rolls and calling the difference inter-rater variation.
If one shift records a borderline pucker as a major defect, another records it as minor and a third accepts it, the quality log contains three decisions but not three physical truths. The resulting Pareto chart can misdirect the next review towards whichever classification habit is most common. This is a measurement-system problem before it is a production problem. The fabric review identifies inconsistency and subjectivity as manual-inspection limitations, and the seam-puckering study was designed specifically to improve grading consistency.[S01][S21]
The same noise also affects any later vision model trained from manual labels. A model cannot learn a stable boundary from labels that change with the observer unless those disagreements are retained and adjudicated. GFT's practical requirement is to preserve the image or physical sample reference, the original inspector decision and the reviewer decision. The reviewer should not silently overwrite the first record. Agreement, escalation and final disposition are different fields.
2.7
Read-out
Illustrative GFT arithmetic, not observed factory data. A 1,000 kg comparison input is split into two independent 500 kg scenarios. Manual: 500 x 72% = 360 packed, 500 x 20% = 100 offcuts, 500 x 8% = 40 quality rejects. Vision-assisted: 500 x 78% = 390 packed, 500 x 18% = 90 offcuts, 500 x 4% = 20 quality rejects. All shares are assumptions. Cited sources support defect detection and resource-efficiency context, not these quantities.
Unit: kg of fabric per illustrative scenario. Status: illustrative. Sources: R29, S30, S32. Illustrative figures are teaching examples, not measured factory results.
Read this exhibit in its chapter ↗2.7 · Introduction
The four-point system gives a fabric buyer and seller a compact way to describe roll quality.
Exhibit 61, chapter 13: Manual input to Packed fabric
2.8 · Penalty brackets convert length into points, not into a complete quality judgement
The penalty schedule in the opened study assigns one point for a fault of less than 3 inches, two points for 3 to 6 inches, three points for 6 to 9 inches and four points for more than 9 inches.
Exhibit 61, chapter 13: Manual input to Packed fabric
2.9 · Manual under-counting creates a false-acceptance band
Let r_points be the share of true penalty points that reaches the tally sheet.
Exhibit 61, chapter 13: Manual input to Cutting offcuts
2.7 · the bodyThe film above follows these sub-steps.
The four-point system gives a fabric buyer and seller a compact way to describe roll quality. The knitted-fabric study used here describes it under ASTM D5430-93 and records its use in quality control, fabric purchasing and buyer-seller agreements.[S0203] Its strength is arithmetic consistency. Its weakness is that the calculation can only process defects that were detected and measured.
The penalty schedule in the opened study assigns one point for a fault of less than 3 inches, two points for 3 to 6 inches, three points for 6 to 9 inches and four points for more than 9 inches. Holes and setups receive four points.[S0203] In millimetres, the corresponding boundaries are 76.2, 152.4 and 228.6 millimetres, derived using 25.4 millimetres per inch. These are length brackets, not a full assessment of colour, handle, dimensions or garment performance.
After the inspector sums the points, the score is normalised to 100 square yards. Using the formula presented in the study, P_100 = (P_observed x 100 x 36) / (L_yd x W_in), where P_observed is total observed penalty points, L_yd is roll length in linear yards and W_in is cuttable width in inches.[S0203] The factor of 36 converts width in inches to yards. The denominator, L_yd x W_in / 36, is roll area in square yards.
Acceptance must then be tied to a stated specification. The opened knitted-fabric study uses Grade A for 1 to 20 points per 100 square yards, Grade B for 21 to 28 and not accepted above 28.[S0203] A factory or buyer may use another threshold, but that threshold is a contractual choice, not a universal constant. This distinction matters because the commonly quoted 40-point rule cannot be treated as though it were the score table of every fabric programme.
The opened knitted-fabric study uses Grade A for 1 to 20 points per 100 square yards
S0203
Let r_points be the share of true penalty points that reaches the tally sheet. This is an illustrative modelling variable, not the same measure as the percentage of individual defects detected. A long defect can carry more points than a short one, so defect recall and point recovery need not be equal. Under a proportional under-counting assumption, P_100,observed = r_points x P_100,true. If the specified rejection threshold is T, a truly rejectable roll can be accepted when P_100,true > T but r_points x P_100,true <= T.
In this illustrative model, the false-acceptance band is therefore T < P_100,true <= T / r_points. Lower point recovery widens that band. The model is deliberately simple: it assumes that missed observations reduce the total proportionally and that no false alarms add points. A real audit should record the actual defects missed, their lengths and any points added incorrectly. Those records can replace the assumption with a measured roll-level error distribution.
the false-acceptance band is therefore T < P_100,true <= T / r_points
illustrative
The example shows why improving arithmetic cannot rescue incomplete observation. It also shows why a roll close to the threshold deserves a second inspection even when the first score says pass. A camera-based implementation can calibrate pixels to inches and assign four-point scores from detected fault dimensions, as demonstrated in the knitted-fabric study.[S0203] Yet automation still requires validation of bounding-box dimensions and missed defects. A digital score is not automatically a true score.
A second limitation is spatial. A total penalty score says nothing about where the defects lie. If the cutting room receives only a total, it cannot place pattern pieces around a local hole or stain without finding it again. The automated four-point prototype links calibrated image coordinates to fault dimensions and scoring.[S0203] That connection between score and location is operationally more useful than a higher total by itself.
Interlude · the cloth half-woven
Woven: sections 2.0 to 2.7Still to weave: 2.10 to 2.15
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 ↗2.10
Acceptable Quality Limit sampling answers a different question from inline inspection. It estimates whether a lot is acceptable from a sample, under a buyer-selected sampling plan. The garment study used here defines AQL as the maximum percent defective, or defects per hundred units, that can be considered satisfactory as a process average for sampling inspection.[S0202] It does not say that every lot at that percentage will pass, nor that the sample percentage is a permit to ship the same percentage of defective garments.
The distinction between acceptance number and AQL is essential. For code letter K in the study's normal, single-sampling example, the sample size is 125. At AQL 1.5, the acceptance number is 5 and rejection number 6. At AQL 2.5, they are 7 and 8.[S0202] Those integers are sample outcomes. AQL 2.5 does not mean that 2.5 defective items are allowed in every sample, and it does not mean that a lot containing 2.5 percent defectives is certain to be accepted.
For an illustrative 12,000-piece lot, a sample of 125 garments covers 125 / 12,000 x 100 = 1.04 percent of the lot. This is an arithmetic coverage calculation using the sample size from the opened study.[S0202] The remaining garments have not been visually checked by that audit. Random selection is what makes the sample decision meaningful. Selecting convenient cartons from the top of a pallet or from one production period weakens that basis.
Clustered production raises two different problems. First, a true random sample from the whole lot can still miss or under-represent a localised burst because sampling variability remains. Second, an informal sample drawn from only some cartons or time periods can be systematically blind to the burst. The remedy for the second problem is better lot definition and stratified random selection across production time, lines and cartons. The remedy for the first is an appropriate sampling plan and, where the risk requires it, inline screening.
The timing problem remains even with a perfectly random sample. A final audit happens after the sampled garments have been produced and usually after they have been packed. A drifting needle, tension fault or print problem can therefore continue while the lot is being assembled. Acceptance sampling can trigger rejection and containment, but it cannot rewind the process to the first defective piece. The garment AQL study itself presents inspection across finished fabric, cutting, sewing, finishing, pre-final and final stages rather than relying on final inspection alone.[S0202]
If the audit misses the burst, defective garments can ship. If it finds enough defects to reject the lot, the factory must contain, re-inspect or rework product after value has been added. Neither outcome is evidence that sampling has no value. It means AQL is a release control, not a substitute for process control. The buyer's contractual audit and the factory's inline prevention system should coexist.
Exhibit 6 places the two manual protocols beside automated screening. It deliberately keeps sourced rules separate from illustrative factory thresholds and derived coverage. The purpose is to show which decision each method can support.
Read-out
| Four-point length brackets | 1 point below 3 in; 2 at 3 to 6 in; 3 at 6 to 9 in; 4 above 9 in or for holes | Undetected or mismeasured faults never enter the score |
|---|---|---|
| Opened knit-study roll grades | A at 1 to 20; B at 21 to 28; not accepted above 28 points per 100 sq yd | Study-specific acceptance, not every buyer's contract |
| Illustrative 40-point roll gate | True score 48.0 becomes 31.2 at 65 percent point recovery | False acceptance when observed points fall below threshold |
| AQL 1.5 and 2.5, code K | Sample 125; accept at 5 or 7; reject at 6 or 8 respectively | Lot release after production, not immediate station correction |
| Illustrative AQL lot coverage | 125 pieces from 12,000 equals 1.04 percent | Unsampled pieces and clustering between sample points |
| Automated fabric and garment screening | Above 90 percent vision benchmark; up to 60 m/min; 0.45 s/image garment result | Local recall, optics, false alarms and physical handling still need validation |
Four-point brackets, study acceptance grades and formula come from S0203. The 40-point row is an illustrative buyer specification used in the chapter case, not a universal ASTM threshold. AQL code K sample and acceptance numbers come from S0202. Coverage of 1.04 percent is derived as 125 divided by 12,000 times 100 for an illustrative lot. Vision benchmarks above 90 percent and up to 60 metres a minute come from R29. The 0.45-second garment image-processing result comes from S19 and excludes physical handling.
Unit: protocol rules and derived coverage. Status: derived. Sources: R29, S19, S0202, S0203.
The table should not be read as a claim that 100 percent automated inspection eliminates error. The inherited register itself reports automated detection above 90 percent, not perfection.[R29] The improvement is continuous screening with structured exception records. AQL can then perform its proper role as an independent lot-release check rather than carrying the whole quality burden.
2.13
Roll-entry interception
Roll-entry interception: total 250. Running total 250 illustrative CFA francs (XAF) lost per defect · illustrative
The cost of a defect depends on where it is stopped.
Cutting loss
Cutting loss: adds 400. Running total 650 illustrative CFA francs (XAF) lost per defect · illustrative
Sewing and rework
Sewing and rework: adds 750. Running total 1,400 illustrative CFA francs (XAF) lost per defect · illustrative
Finishing loss
Finishing loss: adds 1,100. Running total 2,500 illustrative CFA francs (XAF) lost per defect · illustrative
Finished defect
Finished defect: total 2,500. Running total 2,500 illustrative CFA francs (XAF) lost per defect · illustrative
Lot and buyer consequences
Lot and buyer consequences: adds 1,000. Running total 3,500 illustrative CFA francs (XAF) lost per defect · illustrative
Escaped defect
Escaped defect: total 3,500. Running total 3,500 illustrative CFA francs (XAF) lost per defect · illustrative
Illustrative model, not a cost benchmark. Assumed entry interception cost is XAF 250. Assumed additions are XAF 400 after cutting, XAF 750 after sewing and XAF 1,100 after finishing, producing XAF 2,500 or ten times entry cost. A further illustrative XAF 1,000 lot-handling and commercial allocation produces XAF 3,500. S0203 supports the mechanism of unrecoverable processing losses when faults pass early inspection. S08 supports the higher cost assigned to false negatives. No GFT operating cost is claimed.
Unit: illustrative CFA francs (XAF) lost per defect. Status: illustrative. Sources: S08, S0203. Illustrative figures are teaching examples, not measured factory results.
2.13 · the bodyExhibit 07 above; its figures are illustrative where marked.
The cost of a defect depends on where it is stopped. The knitted-fabric study explains that a fault passing inspection can be dyed and then rejected after finishing, consuming time, chemicals, dyestuff and utilities without a recoverable product.[S0203] The same logic continues through cutting and garment assembly. Every completed operation adds material, machine time, labour or logistics to the flawed item. Early detection avoids those later inputs.
The financial comparison between error types is also asymmetric. The fabric-defect study explicitly treats undetected defects, or false negatives, as more costly than falsely rejecting acceptable fabric, and builds false-negative reduction into its proposed system.[S08] A false positive creates recheck work and possible unnecessary rejection. A false negative can create a finished defective garment or a customer escape. This does not make false alarms free. It means the inspection threshold should reflect both costs rather than maximising reported accuracy alone.
Exhibit 7 converts that mechanism into a transparent illustrative case. It starts with an assumed XAF 250 cost to isolate and splice around a fault at roll entry. Cutting, sewing and finishing then add assumed losses until the finished defect costs XAF 2,500, ten times the entry cost. A further assumed XAF 1,000 for lot-level handling and commercial consequences brings the escaped case to XAF 3,500. These are scenario values, not observed Cameroon costs or a general tenfold law.
The model should be replaced with local accounts before any investment decision. Fabric consumed, replacement-panel labour, unpicking time, restitching, pressing energy, downgrade value, reinspection and buyer charges each need a measured basis. Some defects are repairable, while others destroy the garment. Some buyer consequences fall on a whole lot rather than one item. Allocating that lot total to one escaped defect requires an explicit denominator. The exhibit is a method prompt, not a payback forecast.
2.14
Manual inspection consumes time even when it finds the right defects. The fabric review lists time consumption among its main drawbacks, while the Ethiopian garment study describes manual visual inspection as subjective, fatigue-prone and a limit on throughput.[S03][S19] A queue at final inspection also delays information. By the time the tally reaches the sewing supervisor, the line may have repeated the same fault.
The Ethiopian study provides a useful but bounded benchmark. Its VGGNet-19 and artificial-neural-network system used a dataset of 1,000 trouser images collected in Addis Ababa garment facilities under factory lighting. It achieved 86.7 percent overall classification accuracy and processed each image in 0.45 seconds on consumer-grade hardware, reported as a 60-fold improvement over manual inspection throughput.[S19] The result concerns image classification, not complete physical handling, measurement or buyer acceptance.
Read-out
Derived from the opened Ethiopian garment study (S19): reported automated processing is 0.45 seconds per sample and 60-fold manual-equivalent throughput. Manual equivalent is 0.45 times 60 equals 27.0 seconds. The comparison excludes garment placement, turning, dimensional checks, exception handling and packing, so it is not a complete station cycle-time claim.
Unit: seconds per garment sample. Status: derived. Sources: S19.
A line-balancing example shows the denominator. In the paper's illustrative model of 3.0 million garments a year over 300 working days, daily output is 10,000 garments. At 27.0 seconds each, manual inspection requires 270,000 inspector-seconds a day. One eight-hour shift contains 28,800 seconds, so the theoretical requirement is 270,000 / 28,800 = 9.375 full-time equivalents before breaks, fatigue, handling variation or re-inspection. All inputs in this paragraph are illustrative planning assumptions, not a GFT staffing plan.
In the paper’s illustrative model of 3.0 million garments a year over 300 working days
illustrative
In the same illustrative planning case, 0.45 seconds per image requires 4,500 seconds of inference for the daily volume, or 75 minutes of compute time across the available devices. That arithmetic does not justify a one-person inspection station. Garments must still be presented consistently, exceptions reviewed and defective pieces removed. The correct design use of Exhibit 8 is to size image-processing capacity and identify where manual queues form. Station labour requires a separate measured handling study.
0.45 seconds per image requires 4,500 seconds of inference for the daily volume
illustrative
2.15
A before-and-after comparison is valid only if both periods use compatible definitions. GFT's manual baseline should retain roll and bundle identifiers, fabric family, inspection speed, lighting condition, time on task, defect class, severity, inspector decision and downstream disposition. These are recommended data fields for the Cameroon programme, not claims about an already operating factory. The objective is to distinguish changes in production quality from changes in measurement sensitivity.
The reference review also needs honest limits. A slower second inspection by a senior reviewer is a practical reference, not infallible ground truth. Where possible, disputed items should receive a panel decision or a measurement using a ruler, calibrated image or colour reference. The manual detection estimate should then be reported by defect class. A single pooled recall can hide poor performance on exactly the small or low-contrast defects that matter most.
The baseline is therefore both a quality measure and a warning about the measure itself. Manual logs remain useful for management, provided their uncertainty is visible. Vision should then be introduced as a measured change in search coverage and decision consistency, with humans still reviewing exceptions and ambiguous grades. The next chapter defines the defect classes against which that comparison must be run.
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.
10 sources · 58 citation marks on this pageMost cited: S0205, 13 times
Grade AGrade BGrade CKnot size: citations on this page
All 10 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.
c02-S01 · Grade B
YOLO Object Detection for Real-Time Fabric Defect Inspection in the Textile Industry: A Review of YOLOv1 to YOLOv11
Traditional inspection methods rely heavily on human oversight, which makes them prone to subjectivity, inefficiency, and inconsistency in high-speed manufacturing environments.
2025-04-03. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c02-S03 · Grade B
Fabric Defect Detection Using Computer Vision Techniques: A Comprehensive Review
According to research [3], 60, 75% human accuracy is reported to detect fabric defects and the wastage due to fabric defects leads to the high price of product in market.
2020-11-16. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c02-S08 · Grade C
Fabric Defect Detection Using Deep Learning
Typically, traditional inspection systems exhibit an accuracy range of 60-75%. This paper proposes a solution to reduce costs by introducing a fast and automated defect detection system, complemented by human operator decision-making. The system employs a specialized Convolutional Neural Network (CNN) for defect detection. The training process involved the utilization of over 50 defect types to achieve a well-generalized system. Moreover, given that undetected defects (False Negatives - FNs) are more costly to the company than erroneously classifying a non-defective fabric as defective (false positives), FN reduction methods were incorporated into the proposed system.
2023-05-31. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c02-S19 · Grade B
Development of an automatic garment defect recognition system using a machine learning approach
The system processes images in 0.45 seconds per sample on consumer-grade hardware, representing a substantial 60-fold improvement over manual inspection throughput while maintaining consistent defect detection criteria.
2026-09-06. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c02-S21 · Grade B
Image analysis for seam-puckering evaluation
Until now, the standardized evaluation method for seam-puckering grading is still a visual-based, subjective method.
2016-10-14. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c02-S0202 · Grade B
Comparative Study of the Inspection Parameters for AQL 2.5% and 1.5% in Garments Manufacturing Process
It tells that for code letter k and a 1.5% AQL, sample size is 125 and the acceptance number is 5. The rejection number is stated as 6. In all these single sampling plans, the rejection number is one more than the acceptance number. Similarly, for AQL 2.5%, the acceptance number is 7. The rejection number is stated as 8.
2019-05-16. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c02-S0203 · Grade B
An approach to automatic fault detection in four-point system for knitted fabric with our benchmark dataset Isl-Knit
The penalty points for fabric faults are given based on the length of the faults in inches. The criteria maintained in knit dyeing industries for knit fabric inspection are given in Table 12. Length of defects Penalty points Less than 3 inches 1 Point 3 to 6 inches 2 Points 6 to 9 inches 3 Points More than 9 inches 4 Points Hole and Setup 4 Points
2024-08-24. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c02-S0204 · Grade B
DEVELOPING AN INTELLIGENT AUTOMATION AND REPORTING SYSTEM FOR FABRIC INSPECTION MACHINES
Such quality control operations are based on the attendant employee’s detection of the flaws and the unacceptable areas by visual inspection and scanning the overall fabric structure which is being transported on a brightly illuminated straight plane.
2019-03-26. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c02-S0205 · Grade A
Visual Inspection: A Review of the Literature
The decline in performance is typically complete 20 to 35 minutes into the session, and at least half of the final loss occurs during the initial 15 minutes of the watch (Teichner, 1974).
2012-10-01. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗S03 Mathematical Problems in Engineering (Wiley). 2020-11-16.
S08 International Journal for Research in Applied Science and Engineering Technology. 2023-05-31.
R29 MDPI; industry inspection literature. Publication date unrecorded; inherited source-text check.
S0205 Sandia National Laboratories, United States Department of Energy. 2012-10-01.
S0204 TEKSTIL VE KONFEKSIYON. 2019-03-26.
S0202 Journal of Textile Science & Fashion Technology. 2019-05-16.
S0203 Heliyon (Elsevier). 2024-08-24.
S01 Sensors (MDPI). 2025-04-03.
S21 Textile Research Journal (SAGE Publications). 2016-10-14.
S19 Textile Research Journal (SAGE Publications). 2026-09-06.
Keep the thread · chapter 02 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.