Deep-Learning Visual Inspection in Automotive Plants: Reported Results, Costs and Limits

The use case

Cameras and lighting capture images of parts or vehicles on a line, software judges each item, and a person or machine accepts, reworks or rejects it. Two technologies sit behind this description. Rule-based machine vision applies thresholds, templates and measurements written by engineers, and has been used for decades. Deep-learning inspection learns from labelled example images, or in some methods from images of good parts only².

Labels are unreliable guides. A UVeye executive told a trade magazine that vision technology has always been based on AI, whatever the algorithm¹⁰. This paper counts a result as deep learning only where the source says so, and flags the rest. Predictive maintenance is out of scope, except that Audi says it is exploring it with its weld-inspection data⁹.

Where it is running

BMW Group reports using AI image recognition in series production since 2018 (company-reported). Examples include checking that parts are mounted correctly, comparing model badges with order data at final inspection in Dingolfing, and checking for cracks in the Dingolfing press shop. It gives no number of lines or inspection points, and the page opened carries no date⁷.

At Regensburg, painted bodies have passed through an automated surface inspection since March 2022. BMW calls this inspection long established in the industry. It uses deflectometry, in which striped light patterns reveal paint flaws. The release does not describe the detection as learned; the AI controls the sanding and polishing robots. Up to 1,000 vehicles a day are processed, and staff still treat edges and door gaps and perform the final inspection⁸.

Volvo Cars has used UVeye's Atlas tunnel at Torslanda, Sweden, since February 2020. More than 20 cameras scan each finished vehicle for dents and scratches in 5 to 20 seconds. The vendor says Toyota, Skoda and Daimler use it in a limited way (vendor-reported)¹⁰.

Audi analyses about 1.5 million spot welds on 300 vehicles per shift at Neckarsulm. This replaces random manual ultrasound checks of about 5,000 welds per vehicle (company-reported). It uses ultrasound data, not camera images, and is included because it is the best-documented multi-plant rollout⁹.

Valmet Automotive in Finland operated an AI camera at a windshield station, but a planned door-inspection tunnel was not installed during a 2025 thesis project⁶. No source gave the share of manufacturers, lines or inspection points using deep learning.

What the evidence shows

Company and vendor figures. BMW says its press-shop system no longer raises pseudo-defects, meaning dust or oil mistaken for cracks. It attributes this to training on about 100 images per feature, and says reliability reaches 100% after testing. It gives no baseline, test set, period or false-reject rate (company-reported)⁷.

UVeye says pilots at two German carmakers in 2018–2019 found 10 to 40 percent more anomalies than manual inspection, mainly 0.2 to 0.3 mm paint chips and scratches (vendor-reported). More anomalies is not necessarily more true defects, and no false-alarm rate is given¹⁰. Audi's reported gain is coverage, from sampling to every weld. It publishes no detection or escape rate⁹.

Independent measurement. Aust and colleagues (University of Canterbury, 2022) gave the same photographs of aero-engine blades to 50 industry practitioners and to software (measured in a trial). On piece-part images, humans scored 76.2% accuracy against 48.8% for conventional image processing, which also flagged 27.8% of good blades. On borescope images, humans scored 63.8% against 47.4% for a commercial deep-learning prototype.

The software was far faster (25 ms against 20.7 s per blade) and fully repeatable, while operators matched their own earlier decisions 82.5% of the time. A 3D scanner, which is not AI, found every defect on 26 parts. Samples were small (20 borescope images) and came from maintenance, not production¹.

Benchmarks. One 2023 paper reports 99.86% accuracy on 715 test images of cast pump impellers photographed under stable lighting, which is a laboratory result⁴. On MVTec AD 2, a 2025 benchmark with lighting changes and very small defects, current methods remain below 60% average AU-PRO, a localisation score³.

What it takes: cost, integration, governance

Data. BMW says staff photograph components, mark deviations and train a network overnight from about 100 images per feature, without coding (company-reported). It also notes that cracks occur only rarely⁷. Aust and colleagues cite literature saying several thousand images per defect class are usually needed, and more under variable conditions¹.

Hardware and contracts. The UVeye tunnel is 4.5 m wide and up to 4 m high. It produces about 10 GB of images per car and sends data to on-site and cloud servers. UVeye keeps ownership and charges a flat fee per vehicle inspected, with the price unpublished. Carmakers started one shift at a time (vendor-reported)¹⁰. Audi must retrain its model for differences in weld settings at each further plant⁹.

Cost. No source gave system prices or audited savings. The Valmet thesis built a ten-year return model, but its results are confidential. It excluded annual maintenance for lack of data and took its savings assumption from the Volvo case⁶.

Governance. FDA's February 2026 final guidance sets a risk-based approach to assuring software used in medical-device production under 21 CFR Part 820⁵. The page opened names no AI-vision-specific provision. Audi said no independent standards or certifications for AI applications existed. It therefore worked with the German Association for Quality and two Fraunhofer institutes to make its process audit-proof (company-reported)⁹. No source on workplace-monitoring rules was opened.

Where it falls short

Baselines. The only like-for-like comparison with humans found deep learning less accurate¹. Company claims state no human baseline, defect list or period⁷¹⁰. The human baseline is itself unsettled. A 2023 paper cites roughly 80% accuracy for industrial visual inspection from earlier studies⁴, and Aust measured 63.8% to 84.0% depending on the task¹.

Defect type. In the trial, the AI detected no bends and found nicks and dents at rates about 50% below operators'. It also marked acceptable deposits while missing critical defects¹. BMW's fix for dust and oil needed about 100 images of each confusing condition⁷.

Drift. A 2026 review says models can perform poorly after changes in lighting, camera setup, material batch or geometry². In the trial, results depended on camera perspective and lighting¹. The Valmet thesis assessed lighting variation, reflections, position changes and vibration at its station. Its interviewees reported environmental sensitivity and integration difficulty⁶.

Transfer and retraining. Audi retrains per plant⁹, and BMW says the Regensburg process is being rolled out⁸. No source published results from a second plant.

Human checks. People finish and inspect at Regensburg⁸, focus on flagged welds at Audi⁹, and review results on screen at Volvo¹⁰. No source gave inspector hours saved.

Visibility. The thesis author reports that automotive firms are reluctant to share quality-control information. Only one of three approached carmakers responded⁶. Abandoned pilots are therefore unlikely to be public, and none was found. No production false-reject or scrap figure appeared in any source.

Stage: In early use — why

The hypothesis "In wide use" is not supported. Deep-learning inspection is past the pilot stage. BMW reports series use since 2018 and Volvo has run Atlas since 2020⁷¹⁰. But no source measures breadth, and named systems cover specific inspection points. The best-documented rollouts rely on deflectometry or ultrasound data rather than learned image inspection⁸⁹. Machine vision as a whole is widely used; deep-learning inspection with independently measured results is not. "Pilot with results" fits less well, because production use exists, but its results are unverified.

What to watch next

• Audi announced infrastructure for Brussels, Emden and Ingolstadt⁹. Watch for published results from any plant other than Neckarsulm.
• BMW says the Regensburg process is "currently being rolled out"⁸. Watch for named plants and start dates.
• Valmet Automotive's door-inspection tunnel was not installed as of May 2025⁶. Watch for an installation decision.
• The public MVTec AD 2 evaluation server³ will show whether any method exceeds 60% average AU-PRO.

Questions to ask before adopting

• What were the miss rate and false-reject rate on this line, against which human or rule-based baseline, and who measured them?
• How many labelled examples of each defect exist, and what happens for defects seen only a handful of times?
• Who retrains the model when lighting, suppliers, materials or product variants change, and at what cost per change?
• What share of flagged items still goes to human review, and what do false rejects cost in scrap and rework?
• Does the validation evidence meet our customer and regulatory quality requirements, and who owns the image data and the model?