Computer Vision Development in Munich

Automotive / Premium Vehicle Manufacturing

A Munich-based premium vehicle manufacturer (producing 2.5 million vehicles annually across 15 plants worldwide, with 3 major plants in Bavaria — known for industry-leading paint quality that is a core brand differentiator) needed computer vision to replace and enhance manual paint inspection as production volumes increased and quality requirements tightened. (1) Paint quality standards: the manufacturer's "perceived quality" standard required paint surfaces to be free of defects visible to the human eye under any realistic lighting condition — showroom spotlights, outdoor daylight, low-angle parking garage lighting, and everything in between. This was more stringent than industry standard — where most manufacturers inspected under controlled lighting only. Defect types included dust inclusions (particles trapped in paint layers), orange peel (texture irregularity), runs and sags (paint thickness variations), craters (surface depressions), colour mismatch (between panels), and metallic flop (inconsistent metallic particle orientation). (2) Inspection capacity: 85 trained paint inspectors across 3 Bavarian plants, examining every vehicle under controlled fluorescent lighting. Inspector training required 6-12 months to reach full proficiency. Detection rate: 88% on average, dropping to 78% in the final hours of 8-hour shifts. A single escaped paint defect reaching a customer generated an average warranty cost of EUR 2,800 (refinishing in dealer bodyshop) plus brand perception damage. (3) Colour matching: with 120+ exterior colours (including special-order and individual colours), ensuring consistent colour across body panels assembled on different days, painted on different shifts, and potentially painted at different locations was a significant challenge. Human colour matching relied on experienced inspectors viewing panels side-by-side under controlled lighting — a capability that varied between inspectors and degraded with fatigue. (4) Production speed: paint inspection was the bottleneck in the finishing process — vehicles moved through the inspection tunnel at 2 minutes per vehicle, during which inspectors examined all exterior surfaces. Increasing production speed (to meet demand) required either adding inspection stations (capital and space intensive) or reducing inspection time per vehicle (risking quality). (5) Rework tracking: when defects were found, vehicles were routed to the rework area for repair — sanding, polishing, or partial repainting. Rework consumed 5.8% of finishing capacity and was itself a quality risk (rework could introduce new defects). The manufacturer wanted to reduce rework by preventing defects earlier in the paint process — but needed data on which process steps caused which defect types, information that manual inspection records did not capture with sufficient precision.

We developed and deployed a paint inspection vision system over 14 weeks. (1) Multi-angle inspection tunnel: 42 high-resolution cameras arranged in a tunnel configuration — capturing every exterior surface from multiple angles and under multiple lighting conditions (fluorescent, LED spot, LED diffuse, and UV) as the vehicle passed through at normal production speed (no slowdown required). Multi-lighting: each surface photographed under 6 different lighting conditions in rapid succession — deflectometry imaging revealing surface topology, specular imaging detecting gloss variations, and diffuse imaging revealing colour and texture defects. The multi-lighting approach detected defects that were invisible under any single lighting condition but visible to customers in real-world environments — matching the manufacturer's "any realistic lighting" quality standard. (2) Defect detection models: deep learning models trained on 4.5 million annotated paint surface images from the manufacturer's 3-year quality history. Detection sensitivity: sub-50-micron defects detected (smaller than a human hair diameter — well beyond typical human visual detection capability). Classification: 22 defect categories covering all yield-impacting paint quality issues — each classified by type, severity (cosmetic vs. functional), and location. False positive rate: 2.1% (acceptably low — inspectors quickly confirmed or dismissed AI flags). Detection rate: 97.2% (from 88% human inspection — a 9.2 percentage point improvement that significantly reduced customer-reaching defects). (3) Spectrophotometric colour matching: calibrated colour cameras measuring colour difference (Delta E) between adjacent panels, within panels, and against colour standards. Colour accuracy: detecting colour mismatches of Delta E greater than 0.5 (below the threshold of typical human perception under controlled conditions — but potentially visible to discerning customers under certain lighting). Cross-panel colour coordination: the system tracked colour variation across all panels on each vehicle — flagging vehicles where panel-to-panel variation exceeded thresholds even if each individual panel was within specification. (4) Process feedback: defect location and type data automatically fed back to the paint shop process control system — correlating defects with specific paint booths, robots, material batches, and environmental conditions. Predictive alerts: when defect patterns indicated emerging process issues (increasing dust counts suggesting filtration degradation, orange peel trends suggesting viscosity changes), the system alerted paint shop engineers before the issue produced rejectable defects. Root cause analysis: from 3-5 days (manual investigation) to same-shift (automated correlation between inspection data and process parameters). (5) Rework optimisation: for detected defects, the system recommended the optimal rework method (polish, sand-and-polish, or repaint) based on defect type and severity — standardising rework decisions and reducing over-processing (unnecessary full repaint when polishing would suffice). Rework tracking: automated monitoring of rework outcomes — verifying that repaired areas met quality standards and tracking rework-introduced defects.

Detection rate: from 88% (human) to 97.2% (AI) — meaning 9.2% fewer defective vehicles reaching customers. Customer paint complaints reduced by 68% — from the manufacturer's highest warranty cost category to the second-lowest. Warranty savings: EUR 38M annually (fewer dealership refinishing repairs). Inspection speed: vehicles processed at normal production speed (no bottleneck) — camera tunnel inspection taking 45 seconds vs. 2 minutes for human inspection, enabling higher throughput without additional stations. Colour matching: automated Delta E measurement identifying colour variations invisible to human inspectors — 0.3 Delta E precision. Customer complaints about colour mismatch reduced by 82%. Process improvement: paint process defect feedback loop from 3-5 days to same-shift — enabling same-shift corrective action. Paint shop first-run quality (vehicles passing inspection without rework) improved from 94.2% to 97.8% — a 3.6 percentage point improvement that reduced rework volume significantly. Rework: rework capacity from 5.8% to 3.1% of finishing capacity — freeing 2.7 percentage points of finishing capacity for additional production (worth EUR 45M in annual production value at the Bavarian plants). Rework quality improved through standardised AI-recommended repair methods — rework-introduced defects reduced by 55%. Annual value: EUR 45M (production capacity — reduced rework freeing finishing line capacity), EUR 38M (warranty — reduced customer paint complaints and dealership repairs), EUR 22M (brand value — improved paint quality perception supporting premium pricing), EUR 15M (process efficiency — faster defect feedback enabling same-shift correction and reduced paint material waste), EUR 8M (inspection efficiency — reduced manual inspection requirements and standardised quality assessment) = EUR 128M. Development cost: EUR 680,000. Payback: under 2 days.