A four-phase crawl-walk-run roadmap deploys computer vision in step with factory ramp and proves payback inside the PMS

In the illustrative roadmap, Phase 2 begins only after Phase 1 has produced a usable record of local operating conditions. Its first purpose is to improve material yield. A roll detection carries cross-web and along-web coordinates against a roll identifier. Spreading and cutting then transform that roll into plies, panels and bundles. Each transformation needs a linked identifier and an explicit record of splices, end loss and marker movement. Without that genealogy, an accurate roll coordinate cannot reliably locate a cut panel.

The cutting lead approves the disposition rules during early operation. A severe fault may require removal. A minor fault may be acceptable outside a visible panel zone. A questionable mapping may require physical verification rather than automatic nesting. These rules protect two forms of value at once: they avoid sewing known defects into garments, and they avoid discarding usable cloth through excessive caution. The PMS records the resulting recut, splice, downgrade or accepted use so that material savings can be reconciled rather than inferred from model detections.

The second purpose of the scale phase is to reduce avoidable review work and improve class-specific detection on local materials. In this illustrative design, verified crops from Phase 1 become training and validation candidates, but they require deduplication, label review and separation by roll, lot or production run. Nearby frames from one defect must not leak into both training and test sets. Rare critical classes should remain visible in evaluation even when that makes the reported result less flattering.

The outline's planning milestone of 10,000 verified local crops is illustrative, not a statistical guarantee of readiness. A smaller, well-represented and independently reviewed set may be more useful than a larger collection of repetitive or inconsistent labels. The release decision should depend on class coverage, label quality, the proposed operating threshold and performance on a held-out production period. Every new model version also needs a rollback path and a record of the data and configuration that produced it.

Patterned fabrics require particular caution. The reverse-distillation fabric study addresses diverse colours, textures and defect types across MVTec AD, AITEX, TILDA and a textile-facility dataset. That breadth supports testing distillation as a candidate for domain variation, not assuming that one nominal model will cover every print. A new motif or dye lot should trigger a defined review and, where necessary, re-enrolment before the station returns to routine operation.[S12]

A candidate model should run beside the incumbent on comparable production before replacement. The review compares critical-class recall, false alarms per inspected unit or length, pending volume, processing time and downstream escapes. It also examines whether the proposed gain persists across lines, shifts and material lots. A model that improves the aggregate score while weakening a critical class does not pass. Approval covers a named version at a named threshold, not an indefinite permission to update weights.

Phase 2 is also the earliest sensible point for local demand forecasting in the programme sequence, because the factory now has some sales and order history. The fashion-sector figures in the shared register describe potential from artificial intelligence at scale, not a guaranteed result for this plant. Forecasting should remain a separate acceptance exercise with its own baseline and error measures. It must not borrow credibility from the vision station or divert attention from unresolved quality losses.[R27]

Replication follows the same evidence rule as model updates. A second line receives the proven station design only after its own lighting, mechanics, product mix and reference sample have been checked. The original station's event schema and reaction plan can transfer. Its thresholds and performance cannot be assumed to transfer without verification.

In the illustrative roadmap, Phase 3 addresses defects created between the cut bundle and the finished garment. The factory selects critical seams from verified Pareto evidence, buyer requirements and failure severity. It does not place a camera at every needle by default. A station earns its position when earlier detection can contain a suspect bundle, guide a needle or tension check, and prevent repeated rework without creating an unmanageable stop pattern.

The sewing-model evidence supports a measured trial rather than a broad autonomy claim. In one study, adding Coordinate Attention to the authors' YOLOv8 configuration increased mAP@0.5 by 3.7 percentage points. The complete improved configuration also reduced speed from 48 to 43 frames a second on the reported experimental setup. Its authors acknowledge limited generalisation to complex textured backgrounds. A Cameroonian acceptance run must therefore test both detection and production consequences on its own garments.[S17]

Inline reactions should be proportional to risk. A critical seam failure can require immediate containment. Repeated minor faults may justify a run-rule, a supervisor review or a maintenance check rather than an automatic stop. Every intervention records stopped time, affected bundles and the subsequent verified result. This prevents the quality dashboard from claiming a gain while the production record hides lost availability.

Traceability integration follows the same condition gate. Research on Digital Product Passport readiness describes fragmented upstream data, limited digital systems and methodological complexity as barriers for garment manufacturers. It also notes that textile-specific requirements continue to be defined through later acts and product rules. Phase 3 should therefore make quality and material records exportable and traceable, without claiming that a particular passport format is already sufficient for future compliance.[S27]

Partner-led design experiments may begin only after the production data spine is stable. The shared register describes multi-model pipelines spanning fabric construction, shade, weave and silhouette, but those capabilities do not repair an unreliable inspection record. They remain a separate Tier 3 pilot with their own acceptance criteria. The run phase is complete when added digital scope no longer degrades containment, review workload or data integrity.[R31]

Exhibit 64 converts the roadmap into an evidence contract. The thresholds are deliberately expressed as approved local criteria rather than universal numbers. Before each trial, the sponsor, production lead and quality lead must convert them into measurable values suitable for the product and buyer. They must also state who signs the gate and what evidence would block it.

The gate is not a claim that management integration guarantees quality. In the Central Java garment study, digital quality control and production automation jointly explained 64.5 percent of product-quality variance through a survey-based structural model. That association supports combining technology with management routines, but it does not provide a causal effect or a phase threshold for another factory. Local gate evidence remains necessary.[S24]

Exhibit 65 positions the main deployment risks on two illustrative scales. The coordinates are workshop planning scores, not observed frequencies. Their purpose is to pair every material risk with an engineering control and a management owner before capital is released.

Lint and optical fouling require protected housings, cleaning standard work and image-quality checks. Power and storage interruption require conditioned power, clean shutdown, durable local event identifiers and duplicate-free recovery. Print or dye-lot drift requires a reference check, controlled re-enrolment and a route back to the prior approved configuration. These are station-level controls, but the PMS must record completion and exceptions if they are to remain effective.

False-alarm overload requires class-specific threshold review and enough verifier capacity. Inspector bypass requires usable screens, training, measured agreement and a response policy that does not punish honest escalation. Vendor lock-in requires portable data, documented export, local inference and contractual exit rights. The SQC and FMEA method is useful here because it links risk priority to concrete work standards and equipment checks rather than leaving risk as a colour on a matrix.[S26]

Every gate also needs explicit stop conditions. A critical defect escape, a broken roll-to-bundle chain, a rising unreviewed queue, unexplained denominator loss or an irreconcilable savings claim should hold expansion. The response may be to return to shadow mode, restore the previous model, retrain staff or repair the measurement system. A phase gate is credible only when the factory is willing to use it.