Skip to content
GREEN FUTURETECH.
Explore

Research librarySeeing Every Defect

Chapter 04 of 14 · Part II · Draft web edition

Line-scan cameras and multi-angle LED illumination decide 80 percent of inspection reliability before a neural network runs

No deep-learning architecture can recover a sub-millimetre broken thread that was blurred by motion or washed out by diffuse ambient factory light.

Sections
7
Exhibits
4
Minutes
26
Sources
27
EX 20 · 40 m/min: 0.10 mm pixel (fine thread) 6.7, 0.20 mm pixel (standard weave) 3.3, 0.40 mm pixel (coarse screening) 1.7 (kilohertz (kHz))

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.

The chapter as a draft: 7 sections, 2 exhibit wefts, one shuttle.

Lead exhibit · 19LayersderivedOptical and mechanical stack of a 60-metre-a-minute fabric roll line-scan inspection rig
Edge frame buffer and PMS telemetryVirtual frame tile assembly (…Flat-field and RGB-to-CIELAB…Edge inference and PMS Qualit…Optical sensor and lens assemblyDual 4k or 8k trilinear CMOS…Low-distortion apochromatic m…Positive-pressure air knife v…Synchronised multi-angle LEDilluminationCoaxial brightfield bar for s…Grazing darkfield bar for slu…Transmitted backlight for hol…Web transport and kinematic triggerTensioned guide rollers up to…Quadrature rotary shaft encod…Synchronised LED pulse-train…

Read-outEdge frame buffer and PMS telemetry: Virtual frame tile assembly (1,024 lines), Flat-field and RGB-to-CIELAB calibration, Edge inference and PMS Quality Log output

Method and source

Derived engineering architecture combining line-scan camera acquisition up to 60 metres a minute (R29, S14), modular 800 mm smart visual sensor coverage (S15) and encoder-synchronised LED pulse-train control (S0401).

Unit: system layers. Status: derived. Sources: R29, S14, S15, S0401.

The chapter's warpBar: share of the reading · square: an exhibit

Section 4.0 · Opening · 1 min · no exhibit of its own · one movement Go to section ↓

Overture · 1 min

Computer-vision quality control begins in photonics and mechanics, not in neural network weights.[S14] Across industrial textile and garment inspection deployments, optical acquisition geometry, sensor synchronisation and illumination stability determine roughly 80 percent of illustrative system reliability before the first convolutional layer executes.[S14] When a broken cotton warp end measuring 0.20 millimetres across moves over an inspection roller at 60 metres a minute, an improperly triggered camera smears that thread into the surrounding weave background, and diffuse overhead fluorescent tubes wash out its shadow contrast.[R29] Once spatial phase information or chromatic contrast is lost at the complementary metal-oxide-semiconductor (CMOS) photodiode array, no downstream parameter tuning, attention module or anomaly memory bank can reconstruct the missing signal.[S0403] Engineering a dependable vision system for an illustrative 3.0 million garment a year cut-make-trim plant in Cameroon therefore requires specifying the optical chain first: rotary-encoder-triggered line-scan cameras and Contact Image Sensor (CIS) bars for continuous fabric rolls, multi-angle light-emitting diode (LED) illumination tailored to the surface physics of the 42-class defect taxonomy, environmental sealing against tropical cotton lint, and calibrated RGB-to-CIELAB colourimetry.[R29]

¶ 01 · 3 sources

…and illumination stability determine roughly 80 percent of illustrative system reliability before the first…S14 · illustrative

S14R29S0403

4.13 min · 1 sub-steps · 2 exhibits

Moving fabric webs demand CMOS line-scan cameras synchronised up to 60 metres a minute

Step 1 of 610 m/min

Exhibit 20 · derivedRequired line-scan acquisition rate rises linearly with fabric web speed and optical pixel pitch

02.557.51010 m/min20 m/min30 m/min40 m/min50 m/min60 m/min
02.557.51010 m/min30 m/min50 m/min

10 m/min

10 m/min: 0.10 mm pixel (fine thread) 1.7, 0.20 mm pixel (standard weave) 0.8, 0.40 mm pixel (coarse screening) 0.4 (kilohertz (kHz)) · derived

Inspecting continuous fabric webs on an incoming perch frame or stenter exit presents a kinematic problem that two-dimensional matrix cameras cannot solve economically at industrial speeds.

20 m/min

20 m/min: 0.10 mm pixel (fine thread) 3.3, 0.20 mm pixel (standard weave) 1.7, 0.40 mm pixel (coarse screening) 0.8 (kilohertz (kHz)) · derived

30 m/min

30 m/min: 0.10 mm pixel (fine thread) 5, 0.20 mm pixel (standard weave) 2.5, 0.40 mm pixel (coarse screening) 1.3 (kilohertz (kHz)) · derived

40 m/min

40 m/min: 0.10 mm pixel (fine thread) 6.7, 0.20 mm pixel (standard weave) 3.3, 0.40 mm pixel (coarse screening) 1.7 (kilohertz (kHz)) · derived

50 m/min

50 m/min: 0.10 mm pixel (fine thread) 8.3, 0.20 mm pixel (standard weave) 4.2, 0.40 mm pixel (coarse screening) 2.1 (kilohertz (kHz)) · derived

60 m/min

60 m/min: 0.10 mm pixel (fine thread) 10, 0.20 mm pixel (standard weave) 5, 0.40 mm pixel (coarse screening) 2.5 (kilohertz (kHz)) · derived

Method and source · Exhibit 20

Derived from the kinematic line-scan equation Line Rate (kHz) = Web Speed (mm/s) / (Pixel Pitch (mm) x 1,000) across web speeds from 10 to 60 metres a minute (166.7 to 1,000 mm/s, R29, S14) for 0.10 mm, 0.20 mm and 0.40 mm pixel pitches (S15).

Unit: kilohertz (kHz). Status: derived. Sources: R29, S14, S15.

4.1 · the bodyExhibit 20 above; its figures are illustrative where marked.

Inspecting continuous fabric webs on an incoming perch frame or stenter exit presents a kinematic problem that two-dimensional matrix cameras cannot solve economically at industrial speeds.[S14] In Green Future Tech Limited's Cameroon textile programme, incoming cotton rolls sourced from upstream ginning and weaving partners must be screened at web speeds up to 60 metres a minute (1,000 millimetres a second) so fabric inspection does not bottleneck the cutting room.[R03][R29] At 1,000 millimetres a second, a conventional two-dimensional area-scan camera captures rectangular snapshots separated by variable spatial overlaps.[S14] Mechanical slip, tension adjustments or operator slowdowns on the roll winder alter the overlap distance between consecutive frames, causing a broken pick or oil spot to be double-counted or missed in a dead band, while flood arrays across a 1,600 to 2,000 millimetre width suffer from lens vignetting toward the selvedges.[S14][S15]

¶ 01 · 4 sources

…at web speeds up to 60 metres a minute (1,000 millimetres a second) so…R03, R29

Related · Exhibit 05 · chapter 02

S14R03R29S15

Industrial line-scan CMOS cameras eliminate frame-overlap error and two-dimensional illumination drop-off by collapsing the sensor into a single linear array of 4,096 (4k) or 8,192 (8k) pixels (or three parallel red, green and blue arrays in a trilinear sensor) focused onto a narrow optical strip across the fabric width.[S14] Instead of triggering on an internal clock, the line-scan camera receives hardware trigger pulses from a quadrature rotary shaft encoder coupled to an idler roller riding on the moving cloth.[S14] Each time the fabric advances by one exact pixel pitch (such as 0.20 millimetres), the rotary encoder emits a transistor-transistor logic (TTL) or low-voltage differential signalling (LVDS) pulse that fires one line exposure.[S14] Whether the perch accelerates to 60 metres a minute, creeps at 5 metres a minute during threading, or halts while an operator tags a splice, the warp aspect ratio remains geometrically locked to the weft pixel pitch, assembling fixed-height virtual frames (1,024 or 2,048 lines) continuously in the edge buffer.[R29][S0406]

¶ 02 · 3 sources

Each time the fabric advances by one exact pixel pitch (such as 0.20 millimetres)S14

S14R29S0406

As shown in Exhibit 19, the physical reliability of a high-speed fabric inspection rig rests on four hardware layers: web transport and kinematic triggering at the base, synchronised multi-angle LED illumination around the web plane, sealed optical sensors above the roller, and deterministic frame assembly at the edge.[S14] What Exhibit 19 also highlights is a failure mode often overlooked when retrofitting commercial cameras onto textile frames: frequency synchronisation between the camera line-exposure window and the LED driver current regulation.[S0401] In machine-vision measurement systems built around line-scan cameras, discrepancies between the camera exposure frequency and the lighting control pulse train produce periodic intensity variations that mimic weft-way shade barring or miss-picks.[S0401] Eliminating this artefact requires either constant-current analogue LED drivers with ripple below 1 percent (illustrative engineering specification) or an automated frequency-synchronised strobe controller that fires a deterministic current pulse on every encoder trigger edge.[S0401]

¶ 03 · 2 sources

…LED drivers with ripple below 1 percent (illustrative engineering specification) or an automated frequency-synchronised…S0401 · illustrative

Exhibit 19

S14S0401

4.2 Exposure-time kinematics govern motion blur and photon budgets

4.2

Because a line-scan sensor integrates light over a single row of pixels moving past the optical axis, the line rate $f_{\text{line}}$ (hertz) is governed by the ratio of linear web velocity $v_{\text{web}}$ (millimetres per second) to the projected spatial pixel size $p_{\text{spatial}}$ (millimetres per pixel):[S14]

¶ 04 · 1 source

S14

$$f_{\text{line}} = \frac{v_{\text{web}}}{p_{\text{spatial}}}$$

¶ 05

At the maximum web speed of 60 metres a minute ($v_{\text{web}} = 1,000\text{ mm/s}$), achieving a fine resolution of $0.10\text{ mm/pixel}$ requires an acquisition line rate of $10.0\text{ kHz}$, whereas a standard resolution of $0.20\text{ mm/pixel}$ requires $5.0\text{ kHz}$ and a coarse screening pitch of $0.40\text{ mm/pixel}$ requires $2.5\text{ kHz}$.[R29] The reciprocal of the line rate sets the ceiling on the line period $T_{\text{line}} = 1 / f_{\text{line}}$, which equals $100\text{ microseconds}$ at $10.0\text{ kHz}$ and $200\text{ microseconds}$ at $5.0\text{ kHz}$ (derived arithmetic).[R29][S14] To prevent motion smear along the warp axis from exceeding half a pixel during integration, the active electronic exposure time $t_{\text{exp}}$ must be held between 40 and 70 percent of the line period, leaving 40 to 140 microseconds for the silicon photodiodes to collect photons.[S14]

¶ 06 · 2 sources

At the maximum web speed of 60 metres a minute ($v_{\text{web}} = 1,000\text{ mm/s}$)R29

R29S14

Exhibit 20 plots this linear scaling across the operating speed envelope from 10 to 60 metres a minute for three spatial pixel pitches.[R29] While industrial CMOS line-scan sensors support line rates of 20 to 80 kilohertz in silicon, the bottleneck shown by Exhibit 20 as speed climbs to 60 metres a minute is photon starvation during sub-100-microsecond exposures.[S14] Inspecting dark indigo denim or navy uniform twill at 60 metres a minute without high-irradiance LED line lights forces operators either to open the lens aperture (collapsing depth of field) or to boost analogue gain (amplifying shot noise and false alarms).[S0403] Camera selection and LED power density must therefore be co-designed around the darkest shade and finest yarn count in the factory's product mix.[S0405]

¶ 07 · 4 sources

…the operating speed envelope from 10 to 60 metres a minute for three spatial…R29

Exhibit 20

R29S14S0403S0405

4.34 min · 1 sub-steps · read with a related exhibit

Spatial resolution arithmetic sizes pixel pitch to single-thread diameter across wide cotton rolls

Determining the right spatial resolution for a textile vision station requires translating the geometry of cotton yarns and sewing threads into Nyquist-Shannon sampling criteria.[S15] In standard woven and knitted apparel fabrics produced from northern Cameroonian cotton fibre (typically ring-spun or open-end yarns in the 20 to 40 English cotton count range, or Ne 20 to Ne 40), individual single-thread diameters sit between 0.15 and 0.30 millimetres.[R03] Under optical sampling theory, detecting a localized discontinuity such as a broken warp end, a missing weft pick, a needle line, or a micro-pinhole requires that the minimum defect feature span at least 2 to 3 pixels across its narrowest dimension.[S15] If the optical pixel pitch is coarser than the yarn diameter, light from a single broken thread is averaged with intact neighbouring threads inside the same pixel well, reducing contrast below the quantization floor.[S23]

¶ 01 · 3 sources

individual single-thread diameters sit between 0.15 and 0.30 millimetresR03

Related · Exhibit 27 · chapter 06

S15R03S23

Empirical evidence from active textile production lines confirms how strongly native optical resolution governs deep-learning recall.[S0403] In the 2026 LUSITANOv2 benchmark study (25,120 native line-scan images with 18,557 bounding boxes under directional illumination), evaluating RT-DETR-L (0.610 mAP@0.5), Faster R-CNN (0.597) and YOLOv12n (0.592) alongside nine one-class anomaly methods showed that retaining native weave detail improved YOLOv12n performance to 0.598 mAP@0.5 at the highest tested resolution.[S0403] Similarly, the 2026 RAW-FABRID benchmark acquired 709 high-resolution raw fabric images (1,792 by 1,024 pixels) on a custom line-scan inspection machine with controlled LED illumination, providing both full-resolution frames and 256 by 256 pixel patch decompositions to preserve thread-level structure.[S0406]

¶ 02 · 2 sources

evaluating RT-DETR-L (0.610 mAP@0.5)S0403

S0403S0406

4.4 Multi-camera visual arrays against Contact Image Sensor bars

4.4

Covering a commercial fabric roll width of 1,600 to 2,000 millimetres at a thread-resolving pitch of 0.10 to 0.25 millimetres per pixel requires between 6,400 and 20,000 horizontal pixels across the web.[S15] Factory engineers achieve this horizontal pixel density through three optical topologies:[S23]

¶ 03 · 2 sources

…commercial fabric roll width of 1,600 to 2,000 millimetres at a thread-resolving pitch of…S15

S15S23

  • Single 8k or dual 4k overhead line-scan cameras: A single 8,192-pixel CMOS line-scan camera mounted 600 to 900 millimetres above a 1,600 millimetre fabric roll yields a pixel pitch of $0.195\text{ mm/pixel}$, whereas two parallel 4,096-pixel cameras with a 32-pixel overlap achieve the same pitch at half the vertical working distance.[S14]
  • Modular smart visual sensor arrays: In a six-month factory deployment on warp knitting machines, a multi-sensor architecture where each smart visual sensor independently scanned 800 millimetres of web width using wavelet decomposition achieved a 98 percent detection rate on broken-end defects.[S15]
  • Configurable Contact Image Sensor (CIS) bars: A Contact Image Sensor bar integrates linear LED illumination, a one-to-one Graded-Index (GRIN) rod-lens array, and linear CMOS detectors into a sealed bar spanning the fabric width at a 10 to 20 millimetre working distance.[S23] Evaluated on on-loom fabric datasets, a CIS scanning system resolved individual single threads in real time, achieved a zero miss-detection rate on defect-free fabrics, and exceeded 90 percent pixel-level average precision on defective frames.[S23]
List · 3 sources

S14S15S23

Method

Optical sizing equations for a continuous fabric web inspection station

To size the camera array, lens focal length, acquisition rate and edge bandwidth for a fabric roll station of width $W_{\text{web}}$ (millimetres) moving at maximum speed $V_{\text{max}}$ (metres per minute) with minimum defect width $d_{\text{min}}$ (millimetres):[S14]

  1. Spatial pixel pitch ($p_{\text{spatial}}$): Enforce at least two pixels across the smallest target defect: $p_{\text{spatial}} \le d_{\text{min}} / 2$.[S15]
  2. Minimum horizontal pixels ($N_{\text{pix}}$): Add a 5 percent selvedge margin: $N_{\text{pix}} = \lceil 1.05 \times W_{\text{web}} / p_{\text{spatial}} \rceil$.[S14]
  3. Line acquisition rate ($f_{\text{line}}$): Convert speed to millimetres per second ($v_{\text{web}} = V_{\text{max}} \times 1,000 / 60$) and divide by pixel pitch: $f_{\text{line}} = v_{\text{web}} / p_{\text{spatial}}$.[S14]
  4. Lens focal length ($f_{\text{lens}}$): Given sensor width $S_{\text{sensor}}$ and working distance $D_{\text{work}}$: $f_{\text{lens}} = D_{\text{work}} / (1 + 1.05 \times W_{\text{web}} / S_{\text{sensor}})$ (illustrative paraxial equation).
  5. Raw data throughput ($B_{\text{raw}}$): Multiply horizontal pixels by line rate and bytes per pixel $C_{\text{bytes}}$ ($1$ for monochrome, $3$ for RGB): $B_{\text{raw}} = N_{\text{pix}} \times f_{\text{line}} \times C_{\text{bytes}}$.[S14]
  6. Worked example

    Sizing Station A (incoming cotton twill roll perch) for an illustrative 3.0 million garment a year plant

    Consider the incoming fabric perch in Green Future Tech Limited's illustrative 3.0 million garment a year cut-make-trim facility in Cameroon.[R29] The plant receives woven cotton twill rolls of nominal width $W_{\text{web}} = 1,600\text{ mm}$ spun from SODECOTON lint,[R03] and requires catching broken warp ends down to $d_{\text{min}} = 0.40\text{ mm}$ at $V_{\text{max}} = 60\text{ m/min}$ ($v_{\text{web}} = 1,000\text{ mm/s}$).[R29]

    • Step 1 (Pixel pitch): The two-pixel Nyquist rule gives $p_{\text{spatial}} = 0.40\text{ mm} / 2 = 0.20\text{ mm/pixel}$ (illustrative calculation).[S15]
    • Step 2 (Horizontal pixel count): Two synchronised 4,096-pixel (4k) trilinear RGB CMOS cameras ($8,192\text{ active pixels}$, each covering an $800\text{ mm}$ zone) or one 8,192-pixel (8k) camera yields an actual pitch of $1,600\text{ mm} / 8,192 = 0.1953\text{ mm/pixel}$.[S15]
    • Step 3 (Line rate and exposure): At $v_{\text{web}} = 1,000\text{ mm/s}$, the line rate is $f_{\text{line}} = 1,000 / 0.1953 = 5,120\text{ Hz}$ ($5.12\text{ kHz}$), giving a line period of $195.3\text{ microseconds}$ and an exposure window of $t_{\text{exp}} = 100\text{ microseconds}$ (illustrative calculation).[R29][S14]
    • Step 4 (Lens focal length): Two 4k cameras with $7.0\text{ micrometre}$ pixels ($S_{\text{sensor}} = 28.67\text{ mm}$) at $D_{\text{work}} = 700\text{ mm}$ over an $840\text{ mm}$ half-web field require $f_{\text{lens}} = 23.1\text{ mm}$, met by a $24\text{ mm}$ apochromatic lens (illustrative calculation).[S14]
    • Step 5 (Edge data rate): Raw throughput is $41.94\text{ megabytes a second}$ in 8-bit monochrome and $125.83\text{ megabytes a second}$ ($1.007\text{ gigabits a second}$) in 24-bit RGB, carried over GigE Vision or CoaXPress into a local NVIDIA Jetson AGX Xavier or Orin edge node.[S02][S07]

4.53 min · 4 sub-steps · 1 exhibit

Multi-angle LED illumination geometries separate holes, stains and raised three-dimensional defects

Exhibit 216 paragraphs step it

Exhibit 21Control mapillustrativeRelative optical contrast of five canonical textile defect families across four illumination geometries
Transmitted…Coaxial bri…Grazing dar…Cross-polar…Pinholes and reed mar…5222Dye stains and oil sp…1515Raised slubs and knots2253Seam puckering (3D re…1252Satin glare and shade…1325

Read-outPinholes and reed marks: Transmitted backlight 5, Coaxial brightfield 2, Grazing darkfield 2, Cross-polarised RGB 2 (illustrative contrast score (1 to 5))

Method and source

Illustrative optical contrast index from 1 (washed out or invisible) to 5 (high-contrast separation) based on physical light-scattering mechanisms reported in S18, S19, S21 and S23. Scores are an engineering selection guide, not a measured factory trial.

Unit: illustrative contrast score (1 to 5). Status: illustrative. Sources: S18, S19, S21, S23. Illustrative figures are teaching examples, not measured factory results.

4.5 · Introduction · paragraph 1 of 6

Even when a camera lens resolves $0.195\text{ millimetres per pixel}$, a defect only becomes visible to a convolutional kernel if the illumination geometry converts the physical anomaly into a measurable change in pixel intensity or chromaticity.[S0405] Because Chapter 03 established that textile and garment defects span more than 40 classes with different physical morphologies (open voids, flat stains, raised yarn relief, and three-dimensional seam distortion), no single lighting angle reveals every defect family simultaneously.[R29] A high-reliability inspection station therefore combines four optical illumination modes.[S0405]

4.6 Transmitted backlighting for structural voids and density variations

In transmitted backlighting, a diffused LED line light is mounted beneath the fabric web, aimed upward through the cloth into the camera lens axis.[S14] Because intact cotton warp and weft yarns absorb and scatter most incident photons, normal fabric appears as a mid-grey lattice, whereas structural voids (pinholes, tiny holes, broken ends, miss-picks, reed marks, needle lines and thin places) allow light to pass unimpeded, saturating the corresponding pixels.[S18] Backlighting also provides a sharp silhouette at both selvedges for width measurement,[S15] though it remains blind to superficial oil stains, soil spots, or dye discolourations on opaque twills and denims.[S0405]

4.7 Coaxial and diffuse brightfield reflection for chromatic and stain defects

Brightfield front illumination positions linear LED bars on the same side of the fabric as the camera at 45 to 90 degrees relative to the cloth plane.[S0404] Any chemical or pigment anomaly that alters absorption across the visible band (400 to 700 nanometres), including dye spots, oil stains, soil marks, faded prints, misaligned prints and shade drift, produces a sharp reflectance drop or chromatic shift.[S18] In statistical quality control studies of printed T-shirt production where 107 out of 1,380 inspected units failed due to faded prints (43 units), misaligned prints (36 units) and inappropriate colour (28 units), diffuse brightfield RGB illumination isolates all three failure modes.[S26] In the 2026 Addis Ababa trouser factory study evaluating 1,000 garment images across five classes (defect-free, broken stitching, open seams, seam slippage and untrimmed threads), VGGNet-19 plus Artificial Neural Network classification on RGB images achieved 86.7 percent accuracy against 80.0 percent on greyscale images, a 6.7 percentage-point gain that separates stains and off-shade threads from shadow folds.[S19]

4.8 Low-angle grazing darkfield and directional X-illumination for three-dimensional relief

Flat brightfield lighting fails when a raised defect shares the exact colour of the background cloth, such as a cotton slub, snarl, knot, nep cluster, or three-dimensional seam puckering stitched with matching thread.[S18][S21] Grazing darkfield illumination mounts focused linear LED bars at a shallow 5 to 20 degree angle to the fabric plane so flat cloth reflects rays away from the lens while protruding slubs, knots or puckered seams scatter bright light into the aperture and cast sharp leeward shadows.[S21] In automated vision systems for objective seam-puckering evaluation (which with broken threads accounts for 23 percent of sewing defects in Six Sigma jacket studies),[S25] directional X-illumination enables Hough seamline localisation and locally weighted scatterplot smoothing (LOESS) profiling to grade puckering objectively from Grade 5 to Grade 1.[S21]

4.9 Cross-polarised paths for specular glare suppression

When inspecting mercerised cotton sateen, calendered uniform fabrics, or metallic buttons and zips, unpolarised brightfield LEDs generate specular glare off smooth surfaces.[S0405] Specular glare saturates camera pixels to pure white (digital value 255 in 8-bit encoding), masking dye variations and triggering false alarms.[S0405] Placing a linear polarising film over the LED emitter bar and an orthogonal analyser rotated 90 degrees over the camera lens blocks direct specular reflections while transmitting depolarised light scattered from internal dye chromophores.[S0405]

Exhibit 21 synthesises how these four illumination geometries complement one another across five canonical defect families on an illustrative 1-to-5 optical contrast scale.[S18] Transmitted backlighting scores 5 on pinholes and reed marks but drops to 1 on opaque dye stains and seam puckering, whereas grazing darkfield scores 5 on raised slubs and 3D seam puckering[S21] but cannot separate flat chromatic shade drift.[S23] On a line-scan perch frame, engineers achieve multi-modal capture by time-multiplexing odd and even encoder pulses between backlight and brightfield/darkfield LED strobes or by mounting spatially offset lighting zones along the roller path.[S14] What Exhibit 21 cannot replace is physical bench testing on the factory's actual cloth constructions.[S0403]

Key figures from chapter 11 · Exhibit 50Key figuressourcedGarment studies support integrated control rather than isolated detection
  • 0.3Digital quality controlPath coefficient; reported p = 0.002 in S24
  • 0.6Production automationPath coefficient; reported p = 0.000 in S24
  • 64.5%Quality variance explainedJoint result for both predictors in S24
  • 23%Two dominant jacket defectsBroken threads and puckering among 16 types in S25
  • 107 / 1,380Defective T-shirtsObserved units in the S26 case

Read-outDigital quality control: 0.3. Path coefficient; reported p = 0.002 in S24

Method and source

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 ↗

4.102 min · 2 sub-steps · read with a related exhibit

Vision-based tactile sensors and contact image bars isolate structural weave relief from complex print patterns

A persistent vulnerability of standard reflective camera inspection in African apparel manufacturing is the visual complexity of regional fabrics: high-contrast multicolored wax prints (pagne), geometric ndop motifs, and yarn-dyed checks and stripes.[S12] When a brightfield camera images a brightly patterned fabric, the chromatic boundaries of the printed motif generate spatial gradients an order of magnitude stronger than the grey-level disruption of a broken warp thread or weave float.[S0402] While Chapter 05 and Chapter 06 show that Fourier frequency filtering,[S16] Gabor filter banks[S06] and motif-based reverse knowledge distillation[S12] suppress periodic backgrounds in software, pigment-to-structure ambiguity can still mask micro-tears along high-contrast print edges.[S0402]

4.11 Vision-based tactile sensing for colour-blind topography

4.11

When optical separation of structural weave defects from complex surface prints must be absolute, vision-based tactile sensing eliminates pigment interference at the physical sensor boundary.[S0402] Adapted for textile quality control, a vision-based tactile inspection module consists of an enclosed miniature camera, four internal directional LEDs, and a compliant opaque-coated elastic sensing membrane (elastomer) that contacts the fabric surface.[S0402]

¶ 02 · 1 source

S0402

As the fabric slides across the elastic sensing layer, the micro-topography of the warp and weft yarns deforms the compliant elastomer skin.[S0402] Because the outer elastomer coating is optically opaque, the internal camera sees zero photons from the fabric's dye colour, printed motif or ambient lighting.[S0402] Instead, the four internal LEDs illuminate the rear of the deformed membrane, producing a high-resolution relief map of pure physical surface structure.[S0402] Coupled with intensity preprocessing and a Residual Network (ResNet) ensemble classifier, this tactile system detects structural defects across patterned fabrics without external lighting enclosures or recalibration when fabric colours change.[S0402]

¶ 03 · 1 source

S0402

4.12 Contact Image Sensor (CIS) modules for zero-distortion single-thread capture

4.12

Where full-width scanning is required inside confined machinery (such as on a weaving loom or compact verifier), configurable Contact Image Sensor (CIS) bars provide a compact alternative to overhead lens cameras.[S23] Unlike a spherical camera lens projecting a 1,600 millimetre web onto a 28 millimetre chip, a CIS module spans the fabric width using a linear array of Graded-Index (GRIN) rod lenses operating at 1:1 unit magnification alongside integrated linear LEDs.[S23]

¶ 04 · 1 source

Unlike a spherical camera lens projecting a 1,600 millimetre web onto a 28 millimetre chipS23

S23

In a 2024 on-loom fabric defect inspection study, a CIS-based scanning system captured real-time, distortion-free images resolving every single thread across the fabric width.[S23] Paired with a two-level hardware architecture (a low-cost edge node at the loom for anomaly screening and a server for activation-layer-embedded CNN verification), the CIS system achieved a zero miss-detection rate on defect-free fabrics and above 90 percent pixel-level average precision on defective samples.[S23]

¶ 05 · 1 source

…on defect-free fabrics and above 90 percent pixel-level average precision on defective samplesS23

S23

4.133 min · 3 sub-steps · read with a related exhibit

Positive-pressure air knives, sealed housings and rigid frames protect optics in tropical cotton mills

Step 1 of 4Intro

Related exhibit 40 · chapter 08Bar chartderivedGaroua stations require double the LiFePO4 UPS autonomy of Douala-Dibamba rigs to bridge seasonal grid gaps
0200400600260DoualaStation A…600GarouaStation A…150DoualaStation B…345GarouaStation B…110DoualaStation C…255GarouaStation C…

Read-outDouala Station A (Perch): 260 watt-hours (Wh) of LiFePO4 battery capacity per station

Method and source

Derived illustrative battery sizing comparison between the Douala-Dibamba corridor (sized for 2-hour grid/generator transfer and feeder maintenance windows) and Garoua in the northern cotton belt (sized for 4-hour seasonal hydro/feeder outage windows, plus 15% thermal derating and Peltier/air-knife load). Station power baselines follow Exhibit x08-station-power-budget.

Unit: watt-hours (Wh) of LiFePO4 battery capacity per station. Status: derived. Sources: R03, R09, R10, R12, S0801, S0803.

Read this exhibit in its chapter ↗
Related exhibit 47 · chapter 10Before and afterillustrativeA large detection gain does not translate one-for-one into first-pass yield
60708090100Defect captureRework-free completion

Read-outDefect capture: Manual baseline 65, Illustrative vision case 90. Difference +25 (percent, with metric definitions retained separately)

Method and source

Illustrative scenario. Defect capture uses 65 percent as the midpoint of R29's 60 to 70 percent manual range and a conservative 90 percent vision threshold against R29's above-90 summary. Rework-free completion assumes an exclusive 6 percent rework cohort and a 25 percent reduction: baseline 94.0 percent, case 95.5 percent. These are not observed factory results.

Unit: percent, with metric definitions retained separately. Status: illustrative. Sources: R29. Illustrative figures are teaching examples, not measured factory results.

Read this exhibit in its chapter ↗

4.13 · Introduction

Designing an optical system on a clean laboratory bench is very different from keeping that system calibrated across three shifts a day in Cameroon’s industrial zones.

Exhibit 40, chapter 08: Douala Station A (Perch)

4.14 · Laminar air knives and electrostatic lint shielding

Unwinding, brushing, spreading and sewing 100 percent cotton fabric releases cellulose fibrils ("cotton fly" or lint) alongside sizing dust.

Exhibit 40, chapter 08: Douala Station A (Perch)

4.15 · Thermal management and LED lumen maintenance

In high-intensity line-scan illumination, linear LED arrays dissipate significant heat during sub-100-microsecond exposures.

Exhibit 40, chapter 08: Garoua Station A (Perch)

4.16 · Structural vibration isolation on high-speed rollers

At a web speed of 60 metres a minute, eccentric fabric rolls or worn bearings transmit mechanical vibration through the machine frame.

Exhibit 47, chapter 10: Defect capture

4.13 · the bodyThe film above follows these sub-steps.

Designing an optical system on a clean laboratory bench is very different from keeping that system calibrated across three shifts a day in Cameroon's industrial zones.[S0403] Whether a plant is sited in the humid coastal Douala-Dibamba corridor (near the planned Camtext SA hub at Dibamba)[R09][R10] or in the hot northern Garoua cotton belt (near SODECOTON's nine ginning plants, where ambient heat combines with lateritic dust and cotton lint),[R03] three environmental stressors attack unshielded machine-vision hardware: airborne cotton fly, thermal dark-current drift, and roller vibration.[S15]

4.14 Laminar air knives and electrostatic lint shielding

4.14

Unwinding, brushing, spreading and sewing 100 percent cotton fabric releases cellulose fibrils ("cotton fly" or lint) alongside sizing dust.[S15] A lint fibre resting on a glass platen or backlight slot casts a stationary shadow across every subsequent line scan, generating a false vertical warp-break streak down the roll map,[S14] while dust accumulation across an LED diffuser mimics centre-to-selvedge shade variation.[S0405]

¶ 02 · 3 sources

S15S14S0405

To prevent optical fouling without requiring manual wiping every twenty minutes, every GFT fabric inspection station incorporates three physical barriers:[S14]

¶ 03 · 1 source

S14

  1. Positive-pressure laminar air knives: Camera enclosures and under-web backlight troughs use compressed-air manifold slots (fed through a 5-micrometre coalescing filter) to project a laminar curtain of clean, dry air across the optical window, deflecting falling lint (illustrative mechanical specification).[S14]
  2. Ingress-protected IP65 optical housings: Cameras and lenses sit inside anodized aluminium tubes with sapphire or borosilicate front ports sealed by fluorocarbon O-rings (illustrative mechanical specification).[S15]
  3. Antistatic ionising bars at roll entry: Mounting an antistatic bar ahead of the optical slit neutralises static web charge so loose lint is vacuumed away before the imaging line.[S14]
List · 2 sources

S14S15

4.15 Thermal management and LED lumen maintenance

4.15

In high-intensity line-scan illumination, linear LED arrays dissipate significant heat during sub-100-microsecond exposures.[S14] In CMOS silicon, dark-current shot noise doubles for every 6 to 8 degrees Celsius rise in junction temperature, while in white LEDs, elevated junction temperatures accelerate lumen depreciation and shift spectral emission toward blue wavelengths, corrupting CIELAB colourimetry.[S0405] Green Future Tech Limited's optical specification mitigates thermal drift through encoder-gated strobe operation and finned aluminium heat sinks.[S0401] Pulsing the LED bar only during the active 100-microsecond exposure window and holding it off between rolls cuts average thermal dissipation by 40 to 60 percent, extending LED half-life beyond 50,000 hours (illustrative engineering estimate).[S0401]

¶ 04 · 3 sources

linear LED arrays dissipate significant heat during sub-100-microsecond exposuresS14

S14S0405S0401

4.16 Structural vibration isolation on high-speed rollers

4.16

At a web speed of 60 metres a minute,[R29] eccentric fabric rolls or worn bearings transmit mechanical vibration through the machine frame.[S14] If the camera gantry vibrates at 20 to 100 hertz with a lateral amplitude of 0.5 millimetres while sampling at $0.20\text{ mm/pixel}$, straight warp yarns appear as sinusoidal waves, degrading Fourier and Gabor background suppression filters.[S06][S16] The camera bridge and LED bars are therefore mounted on an independent welded steel A-frame anchored to the floor via elastomeric damping pads, decoupled from the roll winders.[S14]

¶ 05 · 4 sources

At a web speed of 60 metres a minuteR29, S14

R29S14S06S16

4.172 min · 1 sub-steps · read with a related exhibit

Trilinear RGB calibration and inline CIELAB delta-E monitoring catch shade drift before cutting

While structural weave flaws damage individual garment panels, unmonitored shade variation across dyed cotton rolls can ruin an entire production lot.[S0405] In cut-make-trim manufacturing, a single shirt, trouser or uniform jacket is assembled from 10 to 25 cut panels.[S25] If a dyed roll exhibits shade drift from left selvedge to right selvedge (side-middle-side shading) or from beginning to end (head-to-tail shading), panels cut from different zones show a visible colour mismatch along the seam.[S0404] As documented in factory statistical quality control studies where inappropriate colour accounted for 28 of 107 defective printed T-shirts,[S26] offline laboratory spectrophotometer spot-checks cannot catch continuous intra-roll shade drift.[S0404]

¶ 01 · 4 sources

trouser or uniform jacket is assembled from 10 to 25 cut panelsS25

Related · Exhibit 36 · chapter 08

S0405S25S0404S26

4.18 Transforming camera RGB into device-independent CIELAB coordinates

4.18

A raw trilinear RGB camera is not natively a spectrophotometer: its red, green and blue values depend on LED spectra and CMOS filter curves.[S0405] As a 2026 review of Industrial Artificial Intelligence methods for fabric colour-defect detection establishes, reliable colour inspection requires colour-consistent representation, illumination stability, and perceptual colour-difference metrics.[S0405] Green Future Tech Limited adopts the online pipeline validated on active textile production lines in 2026:[S0404]

¶ 02 · 2 sources

S0405S0404

  1. Flat-field dark and white shading correction: Prior to threading a batch, the camera captures dark ($I_{\text{dark}}$) and white ($I_{\text{white}}$) reference frames across a polytetrafluoroethylene (PTFE) strip spanning the 1,600 millimetre web width, normalising each pixel via $I_{\text{corr}}(x) = (I_{\text{raw}}(x) - I_{\text{dark}}(x)) / (I_{\text{white}}(x) - I_{\text{dark}}(x))$.[S0404]
  2. Bilateral filtering and K-means colour extraction: An edge-preserving bilateral filter suppresses yarn-crown texture noise while preserving stain boundaries, and K-means clustering extracts the ground shade separately from printed motifs.[S0404]
  3. Calibrated RGB-to-CIELAB transformation: A polynomial matrix calibrated against a 24-patch spectrophotometer chart under D65 daylight converts normalised sRGB values into CIELAB perceptual space ($L^$ lightness from 0 to 100, $a^$ red-green, $b^*$ yellow-blue).[S0404]
  4. Continuous perceptually weighted $\Delta E$ shade mapping: Across nine width zones and every linear metre of roll length, the edge node computes perceptual colour difference $\Delta E$ against the master standard and roll header.[S0404]
List · 1 source

S0404

Evaluated across ten industrial fabric rolls on an active production line, this colourimetry pipeline achieved full roll-level agreement with manual expert inspection while logging continuous longitudinal chromatic profiles.[S0404] Rolls with intra-roll $\Delta E$ below 1.0 are cleared for mixed spreading, rolls between 1.0 and 2.0 are grouped by shade band, and rolls exceeding $\Delta E = 2.0$ are quarantined before cutting (illustrative operational thresholds).[S26][S0404]

¶ 03 · 2 sources

Rolls with intra-roll $\Delta E$ below 1.0 are cleared for mixed spreadingS26, S0404 · illustrative

S0404S26

4.192 min · 1 sub-steps · 1 exhibit

Standard optical bills of materials govern GFT's three shop-floor inspection stations

Exhibit 225 paragraphs step it

Exhibit 22RegistersourcedCamera sensor and illumination selection matrix across fabric roll, cutting table and sewing stations
Station and architecture
Station A1: Dual 4k / 8k CMOS line-scan perch
Coverage or cycle benchmark
Up to 60 m/min web
Primary illumination
Backlight + grazing darkfield LED
Reported performance
Above 90% accuracy across 40+ defects (R29, S14)

Read-outStation and architecture: Station A1: Dual 4k / 8k CMOS line-scan perch. Coverage or cycle benchmark: Up to 60 m/min web. Primary illumination: Backlight + grazing darkfield LED. Reported performance: Above 90% accuracy across 40+ defects (R29, S14)

Open the full register (6 rows)
Station A1: Dual 4k / 8k CMOS line-scan perchUp to 60 m/min webBacklight + grazing darkfield LEDAbove 90% accuracy across 40+ defects (R29, S14)
Station A2: Modular smart visual sensor array800 mm width per sensorHigh-intensity linear LED bar98% broken-end capture on knitting frames (S15)
Station A3: On-loom Contact Image Sensor (CIS)1:1 single-thread pitchIntegrated internal LED barAbove 90% pixel precision, 0% miss on clean cloth (S23)
Station B: Cutting-table area-scan gantryFull marker ply widthDiffuse cross-polarised RGB LED96.3% accuracy, 94.8% F1 on pattern/stitch flaws (S20)
Station C1: Grazing seam-puckering stationStationary seam profileDirectional X-illumination LEDObjective Grade 1 to 5 LOESS/Hough grading (S21)
Station C2: End-of-line garment RGB station0.45 s per garment (60x)Colour-calibrated brightfield RGB86.7% RGB accuracy vs 80.0% greyscale (S19)
Method and source

Compiled from published industrial inspection architectures for fabric rolls (R29, S14, S15, S23), cutting pattern verification (S20) and garment sewing/finishing stations (S19, S21).

Unit: station specifications. Status: sourced. Sources: R29, S14, S15, S19, S20, S21, S23.

4.19 · Introduction · paragraph 1 of 5

To translate optical physics into an executable procurement specification for Green Future Tech Limited's Cameroon textile programme, hardware is standardised around three shop-floor stations: Station A at incoming fabric roll inspection, Station B above the spreading and cutting table, and Station C at inline sewing and end-of-line finished-garment tables.[R29]

Exhibit 22 consolidates the empirical hardware benchmarks across all three factory zones.[R29] Station A requires microsecond line-scan synchronisation at up to 60 metres a minute[R29][S14] or 800-millimetre modular smart sensors[S15] and Contact Image Sensors[S23] to resolve 0.20-millimetre yarns on moving webs, whereas Station B and Station C inspect stationary cut panels and sewn garments using global-shutter RGB area-scan cameras[S19][S20] and directional grazing X-illumination[S21] in 0.45 seconds per piece.[S19] Local freight, customs duties and installation labour in Cameroon are modelled in Chapter 09 and Chapter 10.[R14]

4.20 Station A, Station B and Station C hardware specifications

Mounted on the incoming fabric inspection machine, Station A screens 100 percent of woven and knitted cotton rolls before spreading:[R29]

  • Station A (Incoming fabric roll perch): Two synchronised 4,096-pixel trilinear RGB CMOS line-scan cameras (800 millimetre zone each)[S15] or one 8,192-pixel camera ($0.195\text{ mm/pixel}$ pitch at up to $10\text{ kHz}$ and $60\text{ m/min}$),[R29][S14] or a full-width Contact Image Sensor (CIS) bar;[S23] a 10,000-pulse quadrature rotary encoder with synchronised LED pulse-train controller;[S0401] under-web backlight, coaxial high-CRI brightfield bar for CIELAB $\Delta E$ logging,[S0404] and 12-degree grazing darkfield bars;[S21] plus positive-pressure air knives, antistatic bar and light shroud.[S0405]
  • Station B (Cutting-table overhead gantry): A 12-to-20-megapixel global-shutter RGB area-scan camera with cross-polarised diffuse LED flood panels straddling the spreading table to verify marker alignment and cut-panel geometry, achieving 96.3 percent detection accuracy, 95.1 percent precision, 94.6 percent recall and a 94.8 percent F1-score under the IDAGPD-DL benchmark.[S20]
  • Station C (Inline sewing and end-of-line garment tables): An overhead 5-to-12-megapixel global-shutter RGB camera inside a shrouded tabletop canopy with a colour-calibrated brightfield RGB ring (achieving 86.7 percent accuracy in 0.45 seconds per garment, a 60-fold throughput gain)[S19] and four-quadrant grazing X-illumination for Hough-and-LOESS seam-puckering grading,[S21] complemented on critical lockstitch and chainstitch machines[S22] by a 2-megapixel presser-foot camera feeding YOLOv8 with Coordinate Attention.[S17]

Check the record

The chapter keeps its sources within reach.

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.

27 sources · 195 citation marks on this pageMost cited: S14, 33 times

Grade AGrade BGrade CKnot size: citations on this page

All 27 records shown

R03 · Grade C · inherited record

SODECOTON

Inherited evidence R03

Company profile: more than 220,000 cotton producers supported in northern Cameroon; nine ginning plants; two oil mills.

Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.

Chapter source notes

S14 Applied Sciences (MDPI). 2024-12-23.

R29 MDPI; industry inspection literature. Publication date unrecorded; inherited source-text check.

S0403 Electronics (MDPI). 2026-09-24.

R03 SODECOTON. Publication date unrecorded; inherited source-text check.

S15 Sensors (MDPI). 2013-04-09.

S0406 Data (MDPI). 2026-05-12.

S0401 Production Systems and Information Engineering. 2025-12-09.

S0405 International Journal Science and Technology. 2026-03-01.

S23 University of North Texas Libraries. 2024-05-01.

S02 Sensors (MDPI). 2021-10-25.

S07 arXiv. 2026-08-12.

S18 International Journal of Electrical and Electronics Research. 2024-02-05.

S0404 Scientific Reports (Nature Portfolio). 2026-06-22.

S26 Jurnal Teknik Industri (JATRI). 2026-09-25.

S19 Textile Research Journal. 2026-09-06.

S21 Textile Research Journal. 2016-10-14.

S25 OPSI. 2024-06-30.

S12 arXiv. 2024-01-04.

S0402 arXiv. 2020-03-02.

S16 Textile Research Journal. 2024-09-24.

S06 arXiv / Computer Science and Information Technology. 2022-04-20.

R09 Ecofin Agency. Publication date unrecorded; inherited source-text check.

R10 Ecofin Agency. Publication date unrecorded; inherited source-text check.

S20 International Journal of e-Collaboration. 2026-10-26.

R14 WageIndicator. Publication date unrecorded; inherited source-text check.

S22 arXiv. 2015-10-25.

S17 Journal of Measurements in Engineering. 2024-09-09.

Keep the thread · chapter 04 of 14

Chapter 05 picks up the thread.

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

05Part II: Optical Physics, Imaging Hardware and Deep-Learning Architectures

Modern single-stage and two-stage neural detectors exceed 90 percent accuracy when kernel geometry matches fabric texture

From classical Gabor wavelets to YOLOv11 and Mask R-CNN, architectural choices trade off sub-millimetre recall against edge inference latency.

  • 5.1Classical spectral and statistical filters still earn their place as preprocessing layers
  • 5.3Two-stage instance segmentation earns its compute when the boundary changes the grade
  • 5.5Single-stage detection earns its place through direct, multi-scale prediction
  • 5.7Geometry-specific modules address strip and micro-defects that standard kernels miss
Reading time
23 min
Exhibits
4
Sections
7
Continue to chapter 05
← Previous: chapter 03, Textile and garment defects span more than 40 distinct classes, each with its own optical signature across warp, weft and seamContinue the threadChapter 05 · Modern single-stage and two-stage neural detectors exceed 90 percent accuracy when kernel geometry matches fabric texture

Find a thread.

Type to search.