Computer-vision defect detection is the only industrial AI application a greenfield garment factory can deploy profitably on shift one

A new garment producer enters a market through acceptable product, delivered repeatedly. Computer vision matters commercially because it can make inspection more consistent and its evidence easier to retrieve. The buyer does not purchase an algorithm. The buyer purchases a garment whose material, construction and dimensions meet an agreed specification. The digital system must remain subordinate to that transaction.

The market base case also needs discipline. The inherited evidence records the termination of Cameroon's AGOA eligibility effective 1 January 2020 and lists the country as ineligible for calendar year 2025. On that evidence, the venture should not make duty-free United States access a foundation for its launch economics. Domestic and regional buyers provide the more defensible planning boundary, subject to current trade advice before any commitment.[R18]

Market-size estimates do not remove the need for product-level proof. One commercial research estimate places the African textile market at US$37.3 billion in 2025 and projects US$49.41 billion by 2030. These are vendor estimates for a broad regional category, not measured demand for a Cameroonian entrant or a substitute for buyer discovery.[R24] Quality consistency helps a factory compete for a specified order; it does not create the order.

Domestic operating history offers a warning without supplying a simple cause. The inherited register records CICAM's market share falling from about 85 percent three decades earlier to about 5 percent. It also records July 2023 utilisation of 48.9 percent at CICAM 1 and 10.83 percent at CICAM 2, alongside about CFAF 2 billion of underutilisation cost in 2020. These observations do not prove that inspection failure caused the decline. They show how little room a producer has when capacity and commercial performance are already under pressure.[R11][R12]

For institutional uniforms and repeatable ready-to-wear, a useful quality record answers practical questions. Which roll supplied the panels? Which defects were found? Which units were repaired, downgraded or released? Which model and threshold produced the flag? Who resolved an ambiguous case? Vision can preserve this chain at the point of work. A manual tally entered later can lose the image, location and sequence that make corrective action possible.

The record still has limits. An image of a visible flaw does not establish colourfastness, seam strength, restricted-substance compliance or every dimensional requirement. A stored pass label is not a buyer certificate unless the underlying inspection method and acceptance rule are agreed. The factory should present vision evidence as proof that defined checks occurred, not as proof of properties the station never measured.

Recent African factory research makes the operational possibility tangible. The inherited paper register reports image processing in 0.45 seconds a sample and a 60-fold throughput improvement on consumer-grade hardware in an Addis Ababa trouser study. The same entry records an estimated return on investment of 12 to 18 months. These are study-specific results and estimates, not a supplier quotation, a GFT result or a guaranteed payback for Cameroon.[S19]

The wage context clarifies where the return must come from. Cameroon's monthly minimum wage rose to XAF 43,969 from 23 February 2024, recorded as about US$76 in the inherited register. A business case based only on removing inexpensive inspection labour can therefore be fragile. Avoided fabric waste, reduced rework, fewer downgrades and protected delivery performance must carry the economic argument.[R14]

A greenfield entrant should prefer a tested inspection component over a bespoke research programme, but market growth is only a weak proxy for what can be bought and supported. The inherited market entry values the global fabric-defect-detection AI market at US$523.4 million in 2024 and projects US$6,522.1 million by 2033. It also states 33.8 percent compound annual growth and an Asia-Pacific share of about 46 percent in 2024. These are commercial market-research estimates, not audited shipments or a supplier census.[R28]

The published endpoints and growth rate do not reconcile. GFT arithmetic gives (6,522.1 ÷ 523.4)^(1 ÷ 9) − 1 = 32.35 percent over the nine annual intervals from 2024 to 2033. Applying the stated 33.8 percent instead produces US$7,193.5 million in 2033. Both calculations use the inherited source inputs; neither should be silently adjusted to make the register look consistent.[R28]

Exhibit 3 therefore shows two geometric paths. The endpoint-implied path connects the published start and end values. The stated-rate path preserves the published growth rate and exposes its different endpoint. These are arithmetic interpolations, not annual market observations.

The inconsistency does not invalidate the decision to investigate commercial systems. It does invalidate using the headline forecast as evidence of payback or supplier capacity. Procurement should rely on dated quotations, reference-site evidence, exportable model and data terms, local support arrangements and an acceptance test on the factory's material. A regional market share does not establish that any vendor can maintain a station in Douala or Garoua.

Recent edge research shows both the opportunity and the remaining work. One study deployed a two-stage fabric-inspection cascade on an NVIDIA Jetson Nano, but its disjoint benchmark contained 249 images, of which only 20 were defective. At a recall-prioritised threshold, the first stage detected all 20 defective images while recording a 49.3 percent false-positive rate on non-defective images. The authors positioned the system as assisted triage rather than autonomous acceptance.[S07]

The same study reported 13.45 frames a second for its parallel pipeline against 9.86 for a sequential detector-only loop, while attributing 91 percent of the speedup to overlapping image decoding with inference rather than to the cascade itself. It also reported that 85 percent of forwarded frames were false alarms. These details are more useful to a buyer than a generic real-time claim because they reveal where workflow and computing gains actually arise.[S07]

Buyer adoption is consequently a qualification exercise. The factory should acquire the right to run inference locally, retain its images and annotations, export quality events in an open schema and change integrator without losing its operating records. It should buy demonstrated capability, not a promise that market growth has already solved local inspection.

The argument proceeds through four tests. Data independence asks whether the application can produce a valid local decision without multi-period factory history. Decision proximity asks whether its output acts directly on material or work in progress. Economic mechanism asks which cost is avoided, when it is avoided and who confirms the disposition. Learning curve asks whether each shift creates evidence that improves the next decision. Vision passes these tests conditionally; forecasting and generative design fail the first test at greenfield launch.

These tests also prevent a common sequencing error: purchasing a broad analytics platform before the factory can act on its outputs. A platform may aggregate data efficiently, but aggregation is not an operating decision. The first digital capability should expose a discrepancy that a named role can resolve at a physical boundary. Once that loop is stable, its structured records become dependable inputs for forecasting, optimisation and design rather than another stream of unverified dashboard activity.

The remaining chapters examine the conditions separately. Manual inspection and defect taxonomy establish what must be found. Optics and model architecture determine whether the station can see and classify it. Edge design, procurement and unit economics test whether it can run and pay back. Integration, workforce, traceability and deployment chapters specify how the signal becomes a managed operating capability. None of those later subjects should be used to smuggle an unproven local result into this opening thesis.

The evidence method is correspondingly strict. Shared register entries provide sector context and inherited programme findings. Chapter-local sources support particular research claims. Derived values expose their inputs and operations. Constructed operating examples are labelled illustrative and do not become forecasts through repetition. Where sources use different metrics or disagree, the chapter preserves the difference rather than averaging it away.

Several evidence gaps remain material. Actual product mix, buyer specifications, local defect prevalence, inter-inspector agreement, image quality, review queues, supplier prices, escape costs and operating results are unavailable. The chapter therefore does not claim a realised payback, a validated Cameroonian accuracy or a completed GFT vision deployment. Those questions require local specimens, independent labels, site acceptance tests and dated commercial quotations.

The appropriate investment decision is staged. Approve a bounded inspection cell and its measurement protocol. Test the manual counterfactual and candidate system on the same units. Price each disposition route. Expand only when the recorded net benefit, workforce load and buyer acceptance support expansion. This is how a shift-one application can become profitable without assuming profitability in advance.

Nothing in this chapter constitutes investment, legal or tax advice. Trade treatment, buyer compliance and procurement terms require current specialist confirmation before commitment. The chapter's contribution is a deployment sequence and an evidence standard for that sequence.