Computer Vision Development in Paris

Luxury / Fashion, Leather Goods, and Jewellery

A Paris-based luxury conglomerate (EUR 72B annual revenue, 75 maisons across fashion, leather goods, jewellery, watches, wines, and spirits — 5,500 boutiques worldwide) needed computer vision to combat counterfeiting, automate quality inspection during manufacture, and enable condition assessment for the pre-owned luxury market. (1) Counterfeiting crisis: the conglomerate estimated EUR 3.2B in annual lost revenue from counterfeit products. Super-copies (high-quality counterfeits costing EUR 200-500 to produce, sold at 30-50% of authentic retail price) were increasingly difficult for consumers and even retail staff to distinguish from authentic products. Existing authentication relied on expertise of trained authenticators — but the growing volume of authentication requests (driven by the resale market) exceeded authenticator capacity. (2) Quality inspection fatigue: each maison employed skilled artisans who also performed quality inspection — examining finished products for material defects, stitching irregularities, hardware alignment, and finishing quality. For leather goods, inspectors examined an average of 120 items per day — each requiring 3-5 minutes of detailed visual examination. After 6 hours of inspection, human detection rates dropped by 25% (measured through planted defect studies). Defects escaping inspection and reaching customers generated warranty claims and brand damage. (3) Raw material variability: luxury raw materials (leather hides, exotic skins, silk, cashmere, gemstones) varied significantly in quality — and grading was performed by specialists whose assessments could differ by 10-15% for borderline cases. Inconsistent grading led to premium materials allocated to standard products and standard materials used for premium products — affecting both margin and customer experience. (4) Pre-owned authentication: the luxury resale market was growing at 15% annually — but the conglomerate had limited participation because authenticating pre-owned items required physical inspection by trained experts. The growing volume of authentication requests (85,000 per year, growing 30% annually) exceeded the capacity of the 45-person authentication team. Average authentication time: 25 minutes per item. Queue: 3-4 weeks. (5) Consistency across maisons: each maison had its own quality standards — but enforcement was inconsistent. Visual quality criteria were described in subjective terms ("impeccable stitching," "flawless surface") rather than measurable specifications. Different inspectors at different ateliers applied different standards — creating quality inconsistency across the conglomerate.

We developed and deployed a luxury vision platform over 14 weeks. (1) Authentication AI: multi-modal authentication combining NFC chip verification with visual analysis. Visual markers: deep learning models trained on authenticated product databases (1.2 million verified authentic products across 12 maisons) analysing 200+ visual characteristics — material grain patterns, stitching thread count and consistency, hardware logo engraving depth and precision, edge finishing quality, and overall construction geometry. Micro-feature analysis: high-resolution image analysis detecting features invisible to the naked eye — thread twist direction (authentic products use specific twist patterns), leather grain distribution (genuine leather has natural variation that synthetics cannot perfectly replicate), and hardware plating thickness consistency. Counterfeit detection: the system identified visual indicators that distinguished counterfeits — including super-copies — with 98.7% accuracy (validated against a blind test set of 5,000 authentic and 5,000 confirmed counterfeit items). (2) Quality inspection automation: camera systems installed at quality inspection stations in 8 key ateliers — capturing images of products from multiple angles under controlled, standardised lighting. Defect detection: models trained on each maison's specific quality standards — detecting stitching irregularities, leather surface defects, hardware alignment issues, colour inconsistencies, and finishing imperfections. Severity classification: each detected issue classified by severity (acceptable within tolerance, marginal requiring review, reject requiring rework) — converting subjective quality language into measurable, consistent criteria. Artisan feedback: when defects were detected, the system identified the likely production step and artisan — enabling targeted feedback and training rather than generic quality communications. (3) Raw material grading: visual classification systems for key materials — leather (grade A through D based on surface quality, consistency, and natural markings), exotic skins (pattern regularity, scale quality, and colour consistency), gemstones (4C assessment — cut, clarity, colour, and carat weight estimation from imagery), and textiles (weave consistency, colour uniformity, and defect density). Grading consistency: computer vision grading matched expert consensus 92% of the time (vs. 78% for individual experts) — providing more consistent material allocation across production. (4) Pre-owned condition assessment: standardised condition grading for pre-owned items submitted through boutiques and partner platforms. Guided capture: mobile app guiding the photographer to capture standardised views — overall condition, wear points, hardware close-ups, and interior. Condition scoring: automated 10-point condition score based on visual assessment — accounting for age-appropriate wear, damage, modifications, and structural integrity. Authentication integration: condition assessment combined with authentication — a single process verifying both authenticity and condition. Resale pricing: condition score linked to market data providing estimated resale value — enabling pricing for pre-owned platforms and insurance valuations. (5) Quality standardisation: cross-maison quality metrics — the vision system provided objective, measurable quality data that enabled comparison across maisons and ateliers. Quality dashboard: leadership viewed real-time quality metrics by maison, atelier, product line, and artisan — identifying systemic issues and best practices. Trend monitoring: quality trends tracked over time — detecting gradual degradation (equipment wear, material supply changes) before it reached customer-impacting levels.

Authentication: 98.7% accuracy distinguishing authentic from counterfeit (including super-copies). Authentication processing: from 25 minutes (expert) to 45 seconds (AI + NFC). Authentication capacity: from 85,000/year to 500,000+/year without additional authenticators. Pre-owned market participation: the conglomerate launched its certified pre-owned programme — generating EUR 180M in revenue in the first year (validated by the authentication AI). Quality inspection: defect escape rate reduced by 68% — fewer defective products reaching customers. Inspection speed: from 3-5 minutes per item to 45 seconds per item (with the artisan reviewing AI flags rather than examining every detail). Inspection consistency: coefficient of variation between inspectors reduced from 22% to 5% — standardised quality across all ateliers. Raw material grading: grading consistency from 78% (individual expert agreement with consensus) to 92% (AI agreement with expert consensus). Material allocation optimisation: estimated EUR 25M annual value from better matching material quality to product tier. Pre-owned: condition assessment scaling from 85,000 to projected 500,000+ items per year — enabling the certified pre-owned programme. Customer satisfaction with pre-owned purchases: 4.7/5 (enabled by consistent, accurate condition grading). Annual value: EUR 180M (pre-owned revenue — certified pre-owned programme enabled by authentication AI), EUR 95M (anti-counterfeiting — reduced counterfeit market impact through consumer verification and enforcement data), EUR 42M (quality improvement — reduced warranty claims, fewer returns, and improved customer satisfaction), EUR 25M (material optimisation — better raw material allocation based on consistent grading), EUR 15M (operational efficiency — faster inspection, reduced authentication queue, and standardised processes) = EUR 357M. Development cost: EUR 520,000. Payback: under 13 hours.