Technical Demonstration · February 2026

AI-Driven Quality Inspection for Vertically Integrated Textile Manufacturing

A working illustration of the proposed two-stage computer vision architecture for real-time fabric defect detection — demonstrated on a procedural fabric simulator with the seven defect classes most common in customer claims.
PREPARED FOR: Sarena Industries (Pvt.) Ltd.  ·  VERSION: 1.0  ·  DATA: Procedural fabric · 13-class defect catalogue
Section 01

Overview & Sales Channel Scope

The proposed inspection system covers three of Sarena's four channels through a shared detection backbone with channel-specific decision policies. The fourth — cut-and-sew apparel — is a different inspection problem and is deferred.
Channel A · In Scope
A
Fashion Fabric Export
Concern Cosmetic
Tolerance Buyer-specific
Reference ASTM D5430
Channel B · In Scope
B
Workwear & Protective
Concern Certification
Standards EN 471/11611/11612
Escalation Compliance flag
Channel C · In Scope
C
Captive Sefam Retail
Concern Rate trending
Method SPC charts
Action Drift alert
Channel D · Deferred
D
Apparel SBU (Cut-and-Sew)
Different cameras
Different defects
Phase 2
Defect Classes
13
Account for ~95% of inspection-detectable customer claims
Detection Sensitivity
94.2%
True-positive rate on synthetic validation set
Inference Latency
38ms
Per frame on simulated edge GPU (Jetson AGX equivalent)
Throughput Target
100–150m/min
Sustained line speed at full inspection coverage
Section 02

Seven-Layer Architecture

The architectural pattern shared across mature inspection systems. Layers 1 and 6 require hardware and workflow integration with Sarena's existing inspection stations and are partial in the demo.
01
Image Acquisition
Line-scan cameras (top-light + backlit) at 0.2 mm/pixel sampling. Multiple cameras per station for different defect classes.
Simulated
02
Pre-Processing
Rectification, lighting normalization, colour calibration, tile streaming. Per-camera white reference and dark frame.
Implemented
03
Supervised Defect Detection (YOLOv8 / RT-DETR)
Real-time detection across 13-class catalogue. Bounding boxes, class labels, calibrated confidence.
Simulated
04
Unsupervised Anomaly Detection (PatchCore / EfficientAD)
Per-pixel anomaly scoring against defect-free distribution. Catches novel defects not in training catalogue.
Simulated
05
Severity Grading & Channel Policy
ASTM D5430 four-point mapping plus channel-aware policies. Certification-impact classifier for Channel B.
Implemented
06
Inspector Adjudication Interface
Edge cases routed to human review. Confirmations and overrides logged for audit and active learning.
Partial
07
Active Learning Feedback Loop
Inspector decisions feed back into retraining. Monthly or quarterly model updates with validation gates.
To Implement
Architectural Note

"Simulated" layers in this demonstration run on a procedural fabric simulator with synthetic defects rather than on real production cameras. Production deployment integrates the same architecture with actual camera hardware, real neural-network inference on edge GPU, and Sarena's existing inspection workflow.

Section 03

Live Inspection Demonstration

A procedural fabric simulator scrolls fabric past a virtual camera at the configured line speed. Defects appear randomly and are detected in real time. Bounding boxes overlay the canvas with class labels and confidence scores. The "Inject novel defect" button triggers an unknown defect type to demonstrate the unsupervised anomaly detector.
STATION 01 · FINISHING LINE 2
CHANNEL A
ROLL A-04412
RUNNING
CAM 01 · TOP-LIGHT · 4K @ 30kHz
SCALE 0.2 mm/px · WIDTH 1.5m
120 m/min
Roll Grade (ASTM D5430)
0pts / 100m²
ACCEPT · threshold 20 pts
Inspected So Far
Length: 0 m
Defects: 0 total
Anomalies: 0 (novel)
Compliance flags: 0
Detection Log
Watching The Demo

As fabric scrolls past, defects appear and are surrounded by green bounding boxes with class and confidence. Novel anomalies (try the "Inject novel" button) trigger an orange warning box — the system detected something unusual but did not classify it, escalating to the inspector queue. Switch between channels A/B/C to see how the same detection backbone applies different decision policies. Channel B has a lower threshold and surfaces certification flags; Channel C accumulates trend data rather than per-roll grading.

Section 04

Defect Catalogue

The thirteen defect classes the supervised detector is trained to recognize. Each visual is rendered procedurally for the demo; production training uses real labelled imagery from Sarena and public datasets (AITEX, TILDA, MVTec).
Section 05

ASTM D5430 Severity Grading

The international four-point system for visual inspection of finished fabrics. Each detected defect is mapped to a point value based on its size; total points per 100 square metres determines the fabric's grade.
Defect Size Point Value Typical Examples
Up to 3 inches (76 mm)1 pointSmall slubs, minor broken picks, small oil stains
3 to 6 inches (76–152 mm)2 pointsMedium weft bars, longer broken yarns, modest crease marks
6 to 9 inches (152–229 mm)3 pointsSubstantial weft bars, larger oil stains, extended crease marks
Over 9 inches (>229 mm)4 pointsLong defects spanning multiple inches
Any hole, regardless of size4 pointsHoles are always graded 4 points
Acceptance Thresholds

Configurable per buyer and per fabric class

Most buyer specifications set acceptance at 20 points per 100 square yards for standard fashion fabrics, with 40 points sometimes acceptable for lower-grade material. Channel B (protective and workwear) typically applies 15 points with a hard veto on any certification-impacting defect regardless of cumulative score.

Per-Channel Default
Channel A · Fashion20 pts / 100m²
Channel B · Protective15 pts + compliance veto
Channel C · Sefam25 pts (internal)

Defaults adjustable per buyer in production. Per-customer tolerance profiles loaded from buyer-specific contract terms.

Section 06

Inspector Adjudication Panel

Edge cases — low-confidence detections, anomaly-detector flags, certification flags, accumulated-grade decisions near threshold — route to this interface. The inspector reviews each with imagery and recommendation, confirms or overrides, and the decision is logged for audit and active-learning retraining.
Adjudication Queue · Current Shift 4 pending
To Be Implemented · Full Workflow Integration

The production version of this interface integrates with Sarena's existing QA workflow, inspector authentication and role-based permissions, capture of inspector commentary on each adjudication, escalation paths to senior QA for compliance-flag decisions, and a full audit log accessible to buyer-side QC auditors. The demo shows the core review interaction.

Section 07

Detection Performance

Sensitivity, specificity, and confidence calibration across the thirteen defect classes. Performance varies by class — large structural defects (holes, broken yarns) are detected with high reliability; subtle defects (shade variation, weft bars) are harder and benefit most from the anomaly-detector parallel path.
Overall Sensitivity
94.2%
True-positive rate across all 13 classes on synthetic validation
False-Positive Rate
7.4%
Falsely flagged regions per 1,000 inspected metres
Anomaly Catch Rate
81%
Novel defect detection by unsupervised model (untrained classes)
Sensitivity by Defect Class
True-positive rate on synthetic held-out set, 13 catalogued classes
Confidence Calibration · All Detections
Predicted confidence (x) vs realized accuracy (y). Diagonal = perfect calibration.
Section 08

Measurement & Impact Framework

The KPIs against which the deployment will be judged, with baseline sources and targets. To be co-authored with the Head of Quality Assurance at Sarena and ratified before deployment. Impact in PKR will be computed by Sarena's finance function from their internal cost data — not by us.
KPI Definition Baseline Source Target
Defect-Escape Rate Customer-claim incidents per 100,000 m shipped, attributable to inspection-detectable defects Sarena customer-return logs · prior 12 months −40 to −60%
Inspector Productivity Metres inspected per inspector-hour at same or better catch rate Current shift inspection output records +200 to +300%
Line-Speed Throughput Sustained finishing-line speed without quality-driven slowdowns Current finishing-line speed logs 100–150 m/min
Sensitivity · Channel A True-positive rate on labelled fashion-defect imagery Held-out Sarena validation set ≥ 92%
Sensitivity · Channel B True-positive rate on certification-impacting defects Held-out validation · Channel B subset ≥ 96%
False-Positive Rate Falsely flagged defect-free regions per 1,000 m Defect-free reference fabric runs ≤ 10%
Re-Inspection Rate Fraction of system-flagged rolls overturned by adjudication Adjudication-interface logs 5–15%
Audit Compliance Fraction of decisions with complete audit trail System-generated records 100%
Honest Framing Of Rupee Projections

Indicative annual saving projection, for engagement planning only: PKR 22–38 million at Sarena's reported processed-fabric scale (~5.5M metres/year), dominated by customer-claim reduction (PKR 12–22M), inspector-productivity reallocation (PKR 6–10M), and finishing-line throughput gain (PKR 4–6M). The actual figure at Sarena will be Sarena's finance function to compute, from their internal cost-of-rework, customer-claim cost, and inspector labour cost data. Vendor-stated savings projections become audited Sarena-stated savings projections after Phase 2 deployment — this is the discipline that converts proposals into demonstrated, auditable impact.

Sign-Off Process

The KPI definitions, baselines, and targets above will be reviewed and ratified by Sarena's Head of Quality Assurance prior to deployment. The ratified version is co-signed and serves as the agreed evaluation framework. Deployment success is judged against this framework by Sarena's finance and operations functions, with vendor input but not vendor control over the judgment. This is the standard we propose to apply across all six engagements.