A Proposed Two-Stage Computer Vision Architecture for Real-Time Defect Detection, Severity Grading, and Inspector-in-the-Loop Adjudication at Sarena Industries (Pvt.) Ltd.
We propose a two-stage AI-driven quality inspection system for Sarena Industries' processed woven fabric output, addressing real-time defect detection across three sales channels: global fashion-fabric export (Channel A), workwear and protective fabric export (Channel B), and captive supply to Sefam retail (Channel C). The fourth channel — cut-and-sew finished apparel from the Sarena Apparel SBU — operates on a different inspection problem (stitching, seam integrity, finished-garment dimensional checks) and is deferred to a subsequent engagement. The proposed system combines a supervised real-time detector for the known defect catalogue with an unsupervised anomaly detector for novel defects not represented in training data, mapped onto the ASTM D5430 four-point grading framework and integrated with an inspector-in-the-loop adjudication interface. Channel-aware decision policies handle the distinct concerns of each channel: cosmetic-tolerance variability for fashion, certification-impacting defects for protective fabrics, and rate-trending consistency for repeat constructions. We propose to validate the architecture on synthetic fabric imagery with procedurally injected defects spanning thirteen catalogued defect classes, then calibrate on Sarena's actual inspection imagery once data access is established. Honest expected performance: defect detection sensitivity of 92–96% on the catalogue at false-positive rate below 10%, sustained throughput of 100–150 metres per minute at under 50 milliseconds per frame on edge GPU hardware, defect-escape rate to customer claims reduced by 40–60%, inspector productivity improved 2–3× as inspectors shift from primary detection to adjudication. The system is positioned as a force multiplier for the existing inspection team — not a replacement for human judgment — for credibility, compliance, and audit reasons that we propose to make explicit in all customer-facing communication.
Keywords: computer vision; fabric defect detection; YOLO; unsupervised anomaly detection; PatchCore; ASTM D5430; inspector-in-the-loop; active learning; edge inference; textile quality control; Sarena Industries
We propose to scope the quality inspection engagement to the fabric-mill output across Sarena's three primary B2B fabric channels, parallel to the scope decision made for the Material Requirements Planning engagement. The fourth channel — full-package apparel from the Sarena Apparel SBU — represents a structurally different inspection problem operating on cut-and-sewn garments rather than rolls of fabric, with different defect types (stitching defects, seam integrity, trim placement, finished-garment dimensional accuracy), different lighting and camera setups, and different decision points on the production line. We propose it as Phase 2 work to be addressed in a separate engagement with the Apparel SBU directly.
The architecture we propose uses a single shared detection backbone across the three in-scope channels, with channel-specific policy layers — buyer-tolerance profiles for Channel A, certification-impact mapping for Channel B, statistical-process-control trending for Channel C — applied as configuration rather than code. This mirrors the design pattern proposed for the MRP engagement and preserves the option of unified deployment across channels.
Fabric inspection at Sarena's scale is, at its core, a problem of converting a continuous visual stream — fabric moving past inspection cameras at production speed — into a structured record of defects with class, location, severity, and audit-grade documentation. The conversion must be reliable enough that defects do not escape to customer claims, fast enough that inspection does not bottleneck the line, and discriminating enough that legitimate fabric does not get rejected as defective. The current practice at most mid-sized mills, including Sarena, relies primarily on human inspectors viewing fabric on inspection frames at slower-than-production speeds, with the inspection often serving as the throughput-limiting stage of the finishing line.
We propose to scope the supervised detector to the thirteen defect classes that account for approximately 95% of customer-claim incidents at vertically integrated woven-fabric mills, based on the international defect-classification literature and Sarena's PROBAN and fashion-fabric customer return data. The classes break naturally into three categories: weaving-stage defects (originating at the loom), processing-stage defects (originating in wet processing), and contamination defects (foreign material introduced anywhere in the chain).
Defects falling outside this catalogue — for instance, a process-specific defect that emerges only after a recipe change, or a contamination type never previously observed at Sarena — are the responsibility of the unsupervised anomaly detector described in Section 3, which flags anything unusual without requiring prior class labels. The supervised detector and the anomaly detector run in parallel; the supervised detector handles the common case with high confidence, and the anomaly detector handles the long tail with appropriately wider confidence intervals.
We propose to map detected defects onto the ASTM D5430 four-point system, the international reference for visual inspection of finished fabrics. Under this system, a defect's severity is determined by its size: defects under three inches receive one point, defects three to six inches receive two points, defects six to nine inches receive three points, and defects exceeding nine inches — or holes regardless of size — receive four points. The accumulated points per 100 square yards (or 100 square metres in the metric convention used by European buyers) determines the fabric's grade and acceptance status. Most buyer specifications set the acceptance threshold at 20 or 40 points per 100 square yards, with the specific threshold negotiated per buyer and per fabric class.
The system computes both raw detection (defect found, class identified) and ASTM-mapped severity in a single forward pass, with the point allocation traceable to the underlying detection so that buyer-side QC can audit any grading decision back to the original imagery.
The standard ASTM D5430 framework handles cosmetic defect grading well, but does not adequately address two channel-specific concerns. We propose to handle these through additional decision layers atop the standard grading.
For Channel B (workwear and protective fabric), certification-impacting defects require categorical separation from cosmetic defects. A localized region of a hi-vis fabric where luminance falls below the EN 471 threshold is not merely a cosmetic concern — the fabric fails its certification claim in that region. Similarly, a region of PROBAN-treated fabric where the flame-retardant treatment uniformity is compromised may fail EN 11611 or EN 11612 testing for that lot, with implications far exceeding the cosmetic equivalent. We propose that the system flag certification-impacting regions for separate review and documentation, with explicit reference to the relevant standard.
For Channel C (captive Sefam supply), the inspection concern is less about individual defects per piece (which inevitably occur and are tolerated within standard limits) and more about defect-rate trending across a repeat construction. A construction that has run hundreds of times with a stable defect rate suddenly showing a rate increase signals a process drift — typically a worn loom component, a degraded chemical batch, or an operator change — that warrants investigation before downstream impact accumulates. We propose statistical-process-control charts as a complementary output for this channel, with explicit trend-detection signals surfaced to the QA head.
The HR-version framing of the priority is "AI-powered camera systems identify fabric defects in real-time, reducing the need for manual inspection." The first half of that sentence — automated detection of fabric defects in real time — is what the system delivers. The second half — reducing the need for manual inspection — is true at the level of primary detection of routine defects, but is materially misleading at the level of grade adjudication and buyer-tolerance judgment, which remains and should remain a human task. The system is a force multiplier for the inspection team, not a replacement for it. Sarena's PROBAN business in particular cannot afford a "robots inspect everything" narrative — customers expect human accountability on certified protective fabrics, and the audit trail needs human sign-off at adjudication. We propose to make this distinction explicit in all customer-facing communication.
The proposed architecture follows the seven-layer pattern that has emerged in the fabric-inspection field over the past five years, with reference designs visible in published work from Smartex, the ITMA 2023 inspection systems showcase, the academic computer vision literature on textile defects, and the operational practice of established machine-vision vendors. We propose to follow this pattern at Sarena, with each layer specifically adapted to the channel mix and process speeds described in Section 1.
Line-scan or high-frame-rate area-scan cameras mounted above the inspection frame and finishing line. Recommended configuration for Sarena's fabric width and line speed: line-scan cameras with 4K to 8K horizontal resolution at 25–35 kHz line rate, providing approximately 0.2 mm per pixel sampling at 100–150 metres per minute line speed. Multiple cameras may be configured per inspection station: a primary top-light camera for surface defects, a backlit camera for transparency defects and holes, and optionally a raking-light or polarized camera for specific defect classes that benefit from oblique illumination.
Geometric rectification (correcting for camera angle and lens distortion), lighting normalization (compensating for non-uniform illumination across the camera field), and per-camera colour calibration (using a reference white tile and dark frame). Streaming buffer management to convert the continuous line-scan stream into overlapping tile patches suitable for downstream model inference. Critical for ensuring that detection performance is invariant to lighting drift over a shift and to mechanical adjustments to the inspection frame.
A real-time object detection model trained on the thirteen-class defect catalogue. The proposed family is YOLOv8 or RT-DETR — both established real-time detection architectures with proven sub-50-millisecond inference latency on modest edge GPU hardware (NVIDIA Jetson AGX Orin or comparable). The detector outputs bounding boxes, defect class, and a calibrated confidence score for each detection. Training requires several thousand labelled examples per defect class, sourced from a combination of Sarena's historical inspection imagery, public defect datasets (AITEX, TILDA, MVTec for general patterns), and procedurally augmented synthetic imagery for rare defect classes.
A parallel pipeline that models the distribution of defect-free fabric appearance and flags any region whose appearance deviates from that distribution. The proposed family is PatchCore, PaDiM, or EfficientAD — established unsupervised anomaly detection methods that require only defect-free training imagery and detect anomalies via per-patch feature-distance scoring. The role of this layer is to catch defects the supervised detector cannot — novel defect types, contamination types never previously seen, process-specific defects that emerge from new recipes — at the cost of less specific classification (anomaly score and location, but no defect class).
Mapping of detections onto the ASTM D5430 four-point system, with channel-specific policies applied. For Channel A: per-buyer tolerance profile overlay. For Channel B: certification-impact classifier on top of the detection (does this defect impact the EN 471 luminance claim, the EN 11611 FR claim, the EN 1149 anti-static claim). For Channel C: statistical-process-control trending against historical defect rates for the same construction. The layer produces a per-roll grade, an accept/reject recommendation, a list of cosmetic defects with point values, and (for Channel B) a list of compliance flags with citation to the relevant standard.
The interface through which a human inspector reviews edge cases — detections at low confidence, anomaly-detector flags without specific classification, certification-impact flags, accumulated-grade decisions at the threshold. The inspector confirms or overrides each item, with their decisions logged for audit and fed back into Layer 7. Designed for high-throughput review: a single inspector adjudicating cases from multiple inspection stations simultaneously, with most items requiring less than three seconds of review and only edge cases requiring sustained attention. This layer is where the human-in-the-loop value is realized, and we propose it be designed in close collaboration with Sarena's QA team to match existing inspection workflow rather than imposing a foreign workflow.
Inspector overrides and confirmations from Layer 6 feed back into the supervised model's training pipeline. Cases where the system was wrong (false positives, false negatives, misclassifications) become labelled training examples for the next model iteration. Cases where the system was correct but uncertain (low-confidence detections that the inspector confirmed) are prioritised for inclusion in the next training batch. Anomaly-detector flags that the inspector identifies as a recurring novel defect type can be promoted to a new supervised class. The layer's principal output is monthly or quarterly model retraining with accumulating evidence, with deployment subject to validation gates that prevent regression on the existing performance baseline.
The system is designed so that humans handle the cases the machine handles poorly, and the machine handles the cases the human handles poorly. Humans are good at tolerance judgment, buyer-specific context, certification implications, and novel-defect recognition; the machine is good at sustained-attention scanning at production speed, geometric measurement, statistical-process-control, and audit-grade documentation. The seventh layer ensures the division of labour continues to improve as evidence accumulates.
The fabric-inspection-systems market is characterised by a small number of established machine-vision vendors with long histories in textile inspection, an emerging cohort of AI-native startups bringing modern deep-learning architectures to the problem, and a research literature in academic computer vision that has produced the methodological building blocks now being deployed commercially. We summarize the practice landscape below to anchor the proposed work in observable international context. Specific implementation details inside each vendor's product are largely proprietary; the descriptions are conservative on implementation specifics and accurate at the architectural-pattern level.
Long-established leader in textile testing instruments, with inspection systems including the Q-Bar fabric inspection system and earlier-generation surface defect detectors. Architecture has historically been classical computer vision with rule-based defect classification; recent product generations have incorporated deep learning. Strong installed base across larger Asian and European mills; reference customer for many of the established mid-market players. The Q-Bar architecture is a useful reference for what Sarena's QA team is likely to be familiar with from trade shows and competitor benchmarking.
Specialist in textile finishing-line measurement and inspection systems. Product family includes the Orthopac and Optipac systems for fabric inspection at finishing-line speeds. Architecture combines classical machine vision for dimensional measurement (weft straightness, width control) with deep learning for surface defect detection in newer generations. Strong reputation in European mid-market mills and a credible point of comparison for the Sarena deployment.
One of the most visible AI-native entrants in fabric inspection. Product is positioned as a real-time defect detection system using modern deep learning architectures, deployed at major European fabric manufacturers. Notable for the explicit framing of the system as augmenting human inspectors rather than replacing them — a positioning we propose to adopt at Sarena for the same credibility reasons. Their published case studies report sensitivity in the low-to-mid 90s on common defect classes at production line speeds, consistent with the honest expected performance described in our executive summary.
Broad-spectrum industrial machine vision vendor with textile applications. Architecture is platform-based: the In-Sight and VisionPro toolchains support both classical and deep-learning defect detection workflows. Cognex's position in the textile market is as a flexible platform that a system integrator configures for the specific application, rather than a textile-specific product. Useful as a reference for the platform-versus-product trade-off — the platform approach offers flexibility but requires more integration work; the product approach (Uster, Mahlo, Smartex) offers faster deployment but less customization.
Historically focused on yarn-stage quality monitoring at the spinning machine and winder. Recent expansion into fabric-stage inspection systems. Worth tracking because the upstream yarn-monitoring data could in principle feed into the fabric-stage defect-prediction system, enabling root-cause attribution for yarn-originated defects back to the spinning machine that produced the problematic yarn lot. This integration is rare in current practice but represents a credible future direction.
The takeaway from the survey: the architectural pattern is converging. Two-stage detection (supervised plus unsupervised), ASTM D5430 mapping, inspector-in-the-loop adjudication, and active learning feedback loops are visible across the credible vendors and the academic literature. The differentiation among established vendors sits in installation base, customer-service depth, sensor hardware integration, and the maturity of their training datasets — not in fundamental architectural novelty. The proposed Sarena work positions itself in line with this practice, adapted to Sarena's specific channel mix and integrated with the broader engagement architecture proposed across the six papers.
Computer vision for industrial defect detection requires substantially more data than the structured-data MRP problem. We propose a three-tier data foundation: (a) synthetic fabric imagery with procedurally injected defects for initial methodology development and demonstration, (b) public-domain defect datasets — AITEX Fabric Image Database, TILDA Textile Texture Database, MVTec AD — for supplementary training on common defect patterns, and (c) Sarena's historical and ongoing inspection imagery, calibrated against the production system for the operational deployment.
We propose to generate synthetic fabric imagery procedurally, with explicit control over fabric construction, lighting, and defect injection. The pipeline produces labelled imagery suitable for supervised model training and unsupervised model validation, with the ground-truth defect locations, classes, and severity grades known by construction.
Public-domain defect datasets supplement the synthetic data with real-world textile imagery for which ground-truth labels are available. We propose to use:
The synthetic and public data foundation supports architecture validation and demonstration. Production deployment requires calibration on Sarena's actual inspection imagery, which we propose to address in Phase 2 of the engagement. The calibration phase involves capturing inspection imagery from Sarena's current inspection stations over a defined sampling window, having Sarena's existing inspectors label the captured imagery, training the production models on the labelled Sarena imagery, and validating performance against held-out Sarena-labelled imagery. The synthetic-data foundation accelerates Phase 2 by providing a trained model that requires fine-tuning on Sarena imagery rather than training from scratch.
We propose to be explicit that the synthetic and public-data foundation is for methodology development, architecture validation, and demonstration. Production deployment performance will be determined by the quality and quantity of Sarena's actual inspection imagery and labels, not by the synthetic-data foundation. The synthetic foundation establishes that the architecture is sound and trainable; the real-data calibration establishes that it works for Sarena specifically. Both phases are necessary, in that order.
We propose the following runtime workflow for each fabric roll entering the inspection station. The workflow operates on the trained supervised detector, the trained anomaly detector, the severity-grading layer, and the inspector adjudication interface established during the synthetic and public-data phase, and subsequently fine-tuned on Sarena's imagery during the calibration phase.
At Sarena's typical finishing-line speed of 100–150 metres per minute, with the proposed 0.2 mm per pixel sampling, the system must process approximately 30–50 megapixels per second per camera at sub-50-millisecond inference latency. This is well within the demonstrated capability of YOLOv8 and PatchCore on modern edge GPU hardware (NVIDIA Jetson AGX Orin, current generation), and is the basis for our throughput claim. Sustained operation at this latency is achievable; spikes during edge-case adjudication are absorbed by the inspector's manual review pace rather than blocking the line.
We propose that each of the deployments we deliver to Sarena — this one and the others — be accompanied by an explicit measurement framework defined in advance, co-authored with the relevant department head, and signed off as the basis on which the deployment will be judged. The framework converts methodological promises into audit-grade KPIs and protects against the post-hoc argument about whether the right thing was being measured. For quality inspection specifically, we propose the following framework, to be reviewed and ratified with the Head of Quality Assurance before deployment proceeds.
| KPI | Definition | Baseline Source | Target |
|---|---|---|---|
| Defect-Escape Rate | Customer-claim incidents per 100,000 metres shipped, attributable to inspection-detectable defects | Sarena customer-return logs for preceding 12 months | −40% to −60% |
| Inspector Productivity | Metres inspected per inspector-hour, sustaining the same or better defect-catch rate | Current shift-level inspection output records | +200% to +300% |
| Line-Speed Throughput | Sustained finishing-line speed without quality-driven slowdowns | Current finishing-line speed logs across shifts | 100–150 m/min sustained |
| Sensitivity (Channel A) | True-positive rate on labelled fashion-defect imagery (cosmetic catalogue) | Held-out validation set from Sarena imagery | ≥ 92% |
| Sensitivity (Channel B) | True-positive rate on certification-impacting defect imagery (compliance catalogue) | Held-out validation set, Channel B subset | ≥ 96% |
| False-Positive Rate | Falsely flagged defect-free regions per 1,000 inspected metres | Defect-free reference fabric runs | ≤ 10% |
| Re-Inspection Rate | Fraction of system-flagged rolls overturned by inspector adjudication | Adjudication-interface logs | 5–15% |
| Audit Compliance | Fraction of inspection decisions with complete audit trail (imagery + sign-off + timestamp) | System-generated records | 100% |
The four operational KPIs (defect-escape rate, inspector productivity, line-speed throughput, audit compliance) convert to rupee value via Sarena's own cost data. We propose that the conversion be computed by Sarena's finance function rather than by us, using their internal cost-of-rework, customer-claim cost, inspector labour cost, and finishing-line capacity-cost figures. The framework provides the structured outputs; finance applies the cost factors; the resulting rupee impact is a Sarena-audited number rather than a vendor-claimed number. This is the discipline that converts "demonstrated impact" into "audited demonstrated impact" — a meaningfully stronger position.
Indicative scale, for engagement planning purposes only and to be replaced by Sarena's audited computation: at 5.5 million metres of annual processed-fabric output, with industry-typical customer-claim cost intensities and finishing-line capacity values, the conservative annual saving projection is PKR 22–38 million, 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 to compute and own.
We propose that the KPI definitions, baseline methodologies, and target thresholds in Table 1 be reviewed and ratified by the Head of Quality Assurance at Sarena 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 deployments — the discipline that converts proposals into audited delivery.
We propose a three-phase engagement structure for the quality inspection work at Sarena, parallel in shape to the structure proposed for the MRP engagement.
Complete the architecture specification, implement the synthetic-data foundation, train the supervised detector and the anomaly detector on synthetic and public data, validate calibration and out-of-distribution behaviour, deliver the interactive web demonstration, present the demonstration to Sarena's operational and executive leadership, and co-author the measurement framework (Table 1) with the Head of Quality Assurance for sign-off. Deliverables: this proposal document, the working web demonstration, the architectural specification, the validation report on synthetic and public data, and the signed measurement framework.
Establish camera placement and lighting configuration at one designated inspection station as the pilot. Capture inspection imagery over a sustained window covering a representative cross-section of Sarena's construction mix and defect occurrence. Coordinate inspector labelling of captured imagery. Fine-tune the supervised and anomaly detectors on the Sarena-labelled imagery. Re-validate performance against held-out Sarena imagery. Adjust channel-specific policies to match observed Sarena patterns. Deliverables: Sarena-calibrated detection models, performance validation against Sarena imagery, integration specification for Phase 3.
Deploy the system at one inspection station in advisory mode initially — the system surfaces detections and recommendations, but inspectors retain final authority and the system's outputs do not yet feed into accept/reject decisions. Operate in this mode for a defined evaluation window during which the agreed KPIs are measured against the pre-defined baselines. After demonstrated calibration and impact at the pilot station, expand to additional inspection stations and transition from advisory to integrated decision support. Deliverables: deployed system, operational KPI report against the signed measurement framework, recommendations for full-mill expansion (Phase 4).
We propose to be explicit, with Sarena's leadership and in the engagement documentation, about what the proposed system will not do. It will not replace the inspection team, because grade adjudication and buyer-tolerance judgment remain human tasks. It will not detect every possible defect, because rare and novel defect types will continue to appear and the system will flag them as anomalies rather than classify them. It will not eliminate customer claims, because some fraction of claims relate to inspection-undetectable issues (handle, dimensional stability after wash, perceived quality) that are outside the system's scope. The proposal is for a recommendation engine that delivers calibrated detection with quantified uncertainty, integrated alongside Sarena's inspection team, with measurable impact on the agreed KPIs over the agreed evaluation window. The architecture is established international practice, adapted to Sarena's specific channel mix and integrated with the broader engagement architecture across the six proposed engagements.
Document version 1.0 · Prepared as a working draft for Sarena Industries leadership review · Accompanying interactive web demonstration provided as a separate deliverable · Subsequent versions will incorporate feedback from the Lahore review meeting and the QA-co-authored measurement framework