Opsio - Cloud and AI Solutions
Unbranded silver car body without lights or bumper trim standing on a skid in a white paint-inspection booth, rows of fluorescent tube reflections striped across the bonnet and wing

AI Visual Inspection for Automotive

Automotive paint inspection, weld and casting checks, and final-assembly verification on one platform: catch paint orange peel, weld porosity and missing clips within takt, before they reach rework or your customer.

What is AI visual inspection for automotive?

AI visual inspection for automotive is camera-based defect detection that Opsio installs in your paint shop, body shop, powertrain machining and final assembly, from automotive paint inspection to build checks. Edge deep-learning models find paint surface defects, weld, seam and casting defects, and assembly completeness errors, then tell the PLC to pass, reject or rework the unit.

Key facts

  • Covers paint, weld and seam, casting and machining, and final-assembly defects: 30 named defect types.
  • Paint uses deflectometry: cameras measure distortion in striped light reflected from the paint.
  • Weld bead geometry uses laser-line profilometry (3D); weld surface porosity uses 2D grey-scale imaging.
  • Inference runs on NVIDIA Jetson or Intel OpenVINO edge hardware in under 50 ms per image.
  • Results reach the station PLC over PROFINET, EtherNet/IP or OPC UA within takt.
  • Each result is stored against the VIN or part serial to support IATF 16949 traceability.
  • A feasibility study on your sample parts typically takes 1 to 2 weeks.

Main defect families

Paint surface defects
Faults in the cured paint film, such as orange peel, runs, inclusions, craters and colour mismatch.
Weld, seam and casting defects
Joining, casting and machining faults, such as porosity, undercut, burn-through, cracks, cold shuts and burrs.
Assembly completeness errors
Missing, unseated or wrong-variant clips, fasteners, labels and connectors, checked against the build sheet.

Last reviewed by Praveena Shenoy, Head of AI & Data, Opsio, a Swedish cloud and AI services company. Part of Opsio's AI Visual Inspection service.

Why automotive paint inspection and body-shop checks miss defects

Inspection has to fit inside a station takt of typically 45 to 90 seconds. A body carries several square metres of curved, glossy paint, hundreds of welds and dozens of variant-dependent parts, so one inspector sees only part of it. On a mixed-model line, the next body may differ in colour, trim and options.

Paint is specular: an inclusion a few tenths of a millimetre across shows only at the right angle to a light, and dark metallic colours hide what white solids show. Welds on galvanised steel vary in colour and texture, and machined aluminium reflects unevenly. Rule-based vision tuned on one colour or weld program loses track when the process changes.

Human inspection breaks down late in the shift and at the edges of the body: lower sills, pillars and recesses. An escaped paint defect costs a spot repair or a dealer complaint; a missed weld, crack or unseated connector can become a warranty campaign or a safety recall traced by VIN. Every station the defect passes adds cost.

Machine vision in the automotive industry: three use cases

Each use case covers one plant area: its defects, imaging method and station decision.

Paint and body

Glossy dark grey painted body panel reflecting light-tunnel stripes, with a small dust inclusion and a thin trail in the clear coat and fine orange peel texture
Dust inclusion and orange peel on a painted body panel under light-tunnel stripes

Situation

Your painted bodies leave the topcoat oven on skids and enter a light tunnel. Inspectors walk the body under striped lights, mark defects with tape and route the car to spot repair. Coverage varies by person and shift, and defect positions are rarely recorded well enough to trace a booth, robot or batch. See Opsio's guide to paint defect detection with vision AI.

Defects

  • Orange peela dimpled paint surface that looks like citrus skinLong- and short-wave waviness in the clear coat, caused by poor atomisation, high viscosity or paint that cannot flow out before it cures.
  • Runs and sagsDownward flows on vertical panels, caused by excess film build, low viscosity or slow solvent flash-off.
  • Dust and fibre inclusionsParticles trapped in the film as raised bumps, carried in by booth air, skid debris, clothing or an unclean substrate.
  • Cratersfisheyes: round depressions where wet paint pulls away from a spotThey form when silicone, oil or another low-surface-tension contaminant repels the paint.
  • ScratchesLinear marks in the clear coat from handling, skid contact or sanding, often visible only at certain angles.
  • Colour mismatch between panelsA shade difference between the body and add-on parts such as bumpers or fuel doors, painted on separate lines or batches.
  • Thin spotsAreas of low film build or poor hiding, caused by robot path gaps, electrostatic shadowing in recesses or gun faults.

Inspection setup

The body passes through a deflectometry tunnel: LED screens or bars project striped patterns onto the paint, and 5 to 12 MP GigE Vision area cameras with 16 to 35 mm lenses measure how the reflected stripes bend, typically at 0.1 to 0.2 mm per pixel. Orange peel, runs, inclusions and craters distort the pattern even on dark metallic colours; diffuse channels catch scratches and thin spots.

Colour is never judged from RGB alone: a multi-angle spectrophotometer, or a colour camera calibrated against spectrophotometric reference tiles, reports ΔE (a standard colour-difference value) across panel joints. The station sits after topcoat cure, optionally also after primer. Each defect is logged with body coordinates and the body is flagged for spot repair. The striped-light and line-scan surface methods are shared with coil and sheet inspection.

Outcome

Every body is inspected the same way, with defect type, size and position recorded against the VIN. Repair crews get a defect map instead of tape marks, and paint engineers see which booth, robot or colour produces each defect while the drift is still running.

Body shop and powertrain: welds, seams, castings and machined parts

Macro of a rippled MIG weld bead on a steel plate with several open surface pores in the bead, heat-tint colours along the toes and scattered spatter droplets on the plate
MIG weld bead on steel showing surface porosity and spatter.

Situation

Your body shop joins body-in-white (BIW, the welded sheet-metal shell before paint) in robot cells with MIG/MAG, laser and spot welds, hems and sealer beads; your powertrain plant machines aluminium die-cast housings. Welds are chiselled on a sample basis and castings checked by eye; escapes surface at paint, engine test or in the field. More on automated welding inspection.

Defects

  • Weld porositySurface gas pores from lost shielding gas, zinc vapour or contamination.
  • UndercutA groove melted into the base metal at the weld toe, from excess current or travel speed.
  • Missing or short weldsA weld absent or too short after a robot fault, wire-feed stop or program error.
  • SpatterFused metal droplets beside the weld that disturb fit-up, sealing and paint adhesion.
  • Burn-throughA hole melted through thin sheet where heat input exceeds what the joint can absorb.
  • Misplaced seamsA weld or seam off its nominal path, from fixture wear or fit-up.
  • Incomplete fusionsurface indications onlyFiller lying on the base metal without melting into it at the bead edge; internal fusion still needs NDT.
  • Hem and sealer bead breaksGaps or thin sections in a sealer bead or hem, from dispense-pressure drops or clogged nozzles.
  • CracksCasting fractures from hot tearing during solidification or stress during ejection and trimming.
  • Casting porosityGas or shrinkage voids opened up on machined faces, from trapped air or uneven die cooling.
  • Cold shutsa seam where two metal flows met without fusingThey form when melt fronts meet at too low a temperature in the die.
  • BurrsRaised edges from milling, drilling or trimming that damage seals.
  • ChipsBroken edges or corners from handling or tool breakout.
  • Missing holes or threadsDrilling or tapping steps skipped after a tool breakage or program error.
  • Dimensional deviationsPosition, diameter or flatness out of tolerance, from tool wear, thermal drift or fixturing.
  • Residual swarfMachining chips left in bores, threads or passages after washing, causing leaks.

Inspection setup

A robot-mounted laser-line profiler (laser triangulation, 3D) measures bead width, height, length, position and undercut depth, typically at 0.02 to 0.1 mm height resolution, and follows hem and sealer beads too. A 2D grey-scale camera under a diffuse dome images the bead for porosity and spatter. Castings turn on a rotary fixture under low-angle dark-field LED bars that make cracks, burrs and chips stand out, as in turbine blade crack screening. Structured light checks dimensions; telecentric lenses (constant magnification through depth) gauge holes and threads. Bad welds go to repair, bad castings to quarantine.

Outcome

Every weld and machined face is checked, not a sample. Bead trends per robot and weld program show drift before welds fail, and casting defects are tied to die and shot.

Final assembly

Inner steel door panel on an assembly line with a row of grey plastic trim clips, one clip position empty with only its bare mounting hole showing, a wiring harness and connectors running below
Vehicle door inner panel with one trim clip missing from a row of clips.

Situation

Your trim and final lines fit hundreds of parts per vehicle, with many variants on one line. Operators confirm clips, connectors and labels by sight and sign the build sheet. A missing grommet or unseated connector shows up later as a water leak, a rattle or a warning lamp at the dealer. See assembly line inspection with AI.

Defects

  • Missing clipsRetaining clips for trim, harnesses or pipes not fitted, so parts rattle or come loose.
  • Missing fastenersScrews, nuts or bolts not installed, often after a tool fault or a line stop.
  • Missing labelsAirbag, tyre-pressure or other mandatory labels absent or in the wrong position.
  • Missing grommetsRubber plugs not fitted in body holes, which lets in water and noise.
  • Connectors not seatedElectrical connectors pushed in but not latched, so the circuit opens under vibration.
  • Wrong part variant for the VINA part that fits but belongs to another trim, market or option, picked from the wrong bin.
  • Torque marks absentPaint-pen witness marks missing on safety-critical joints, so you cannot show the joint was torqued and checked.

Inspection setup

Each station has 2 to 8 colour area cameras (5 to 12 MP, GigE Vision) on a fixed gantry or a robot, under diffuse bar or dome lighting that suppresses glare on trim. When the VIN is read from the skid RFID tag, the system loads the expected configuration from the build sheet and checks each region for presence, position and variant; connector seating is judged from latch position, torque marks from colour. A PLC handshake per station returns OK or NOK within takt: NOK holds the vehicle or flags it for the rework bay and raises an Andon call.

Outcome

Every vehicle is checked against its own build sheet, not a generic checklist. Wrong variants and missing parts are caught at the station that fitted them, while the operator can still fix them in cycle, and each check is stored as evidence against the VIN.

How AI inspection runs on an automotive line

Every station follows the same six steps, whether it checks paint, welds, castings or a door.

Grey robot arm holding a black laser-line profile sensor that projects a red line across a weld bead on a body-in-white panel inside a fenced weld cell
Robot-mounted laser-line profilometer scanning a body-in-white weld seam
  1. 1

    Trigger and part presentation

    A skid sensor or conveyor encoder triggers capture as the unit arrives, and the VIN or serial is read from RFID or a DataMatrix code. Bodies move through the tunnel; castings are clamped on a rotary fixture.

  2. 2

    Image acquisition

    GigE Vision area cameras, laser-line profilers and, where needed, line-scan cameras with C-mount or telecentric lenses image the part under deflectometry screens, dark-field bars or dome lights, inside enclosures that block plant light and weld flash.

  3. 3

    Edge inference and latency budget

    Models run on NVIDIA Jetson or Intel OpenVINO edge PCs in the station cabinet, in under 50 ms per image. A paint tunnel takes hundreds of images per body, so the budget is set per body, with margin for the PLC handshake.

  4. 4

    Decision and actuation

    The result goes to the station PLC over PROFINET, EtherNet/IP or OPC UA. The PLC holds the skid, routes the body to repair, diverts a casting to quarantine or releases the unit; MES and Andon get the same event.

  5. 5

    Traceability and data retention

    Each decision is stored with images, defect class, position and model version against the VIN or serial, on the edge and your on-premises server, for as long as your record-control rules require.

  6. 6

    Active learning for new defects and variants

    Uncertain images go to an operator review queue. Confirmed labels feed retraining on SageMaker, which adds new colours, model years, weld programs and variants; each model is validated before it reaches the edge.

Imaging methods for automotive defects

Each defect family has a standard imaging method; the model only finds what the image shows.

Defect familyBest-suited imagingWhy
Orange peel, paint runs and sagsDeflectometry (reflected stripe patterns) in a light tunnelWaviness and flows change surface slope, which bends reflected stripes even where brightness looks uniform.
Inclusions, craters, scratches and thin spotsDeflectometry plus diffuse bright-field area camerasBumps and craters distort reflections; scratches and thin film change contrast under diffuse light.
Colour mismatch between panelsMulti-angle spectrophotometer or colour camera calibrated to spectrophotometric referencesMetallic colours shift with viewing angle; only calibrated measurement gives a repeatable ΔE across joints.
Weld porosity and spatter2D grey-scale camera with diffuse dome lightDome light removes glare from the rippled bead, so pores show dark and droplets bright.
Weld undercut, burn-through, missing or misplaced weldsLaser-line profilometry (3D) on the robot or fixedUndercut depth, bead height, length and position are geometric and need height data.
Hem and sealer bead continuityLaser-line profilometry along the bead pathBreaks and thin sections show as a missing or low bead cross-section.
Casting cracks, casting porosity and cold shutsLow-angle dark-field LED bars on a rotary fixtureRaking light catches crack and pore edges that direct light hides on machined aluminium.
Burrs, chips, swarf, missing holes or threadsDark-field plus backlight or 3D structured lightRaised edges scatter low-angle light; backlight and 3D confirm holes and threads are present and clear.
Dimensional deviations3D structured light or laser triangulation; telecentric lenses for 2D gaugingTelecentric optics keep magnification constant through depth, so diameters and positions measure correctly.
Missing or unseated parts, wrong variant, absent torque marksMulti-camera colour 2D, diffuse light, keyed to the build sheetPresence, position and colour show in 2D; the build sheet says what this VIN should carry.

Integration with MES, Andon and IATF 16949 traceability

The system reads build context from your plant systems and writes every result back against the VIN.

MES and Andon
Reads VIN, build sheet and colour code from MES; writes OK/NOK, defect class and position per station; raises Andon calls on a NOK or a rising defect rate.
QMS and SPC
Sends defect counts per booth, robot, weld program, die and shift to SPC and your QMS, so drift triggers corrective action early.
ERP and supplier quality
Exports defects on bought-in castings and parts with serial and lot for supplier claims and 8D reports.
IATF 16949 traceability
Designed to meet the traceability requirements of IATF 16949: every result, image and model version is linked to the VIN or part serial.
PPAP and inspection capability
Supports PPAP documentation of inspection capability with control-plan entries, attribute agreement analysis (measurement system analysis for pass/fail checks) and acceptance records.
Reject-to-rework loop
Rework stations see the defect map, record the repair and re-inspect, so a repaired unit closes its own record.
Where data lives
Images and decisions stay on the edge PC and your on-premises server; retraining data can sit in an EU cloud region. You own your data.

From first test to your line in five steps

You test the method on your own parts before committing, and the station is proven in the AI Lab before it reaches your line.

  1. 1

    Opsio runs tests in the AI Lab

    You ship a sample set: typically 20 to 50 good parts and 10 to 30 examples per defect class, such as painted panels with orange peel, runs, inclusions and craters plus colour references; welded sub-assemblies with porosity, undercut and burn-through; castings with cracks, porosity and burrs; or a door module with missing parts. Opsio images them with candidate cameras and lighting, trains a first model and measures detection and false-reject rates per defect class.

    You provide
    Sample parts per defect class, defect catalogue, boundary samples.
    Opsio delivers
    Imaging trials, first model, per-defect results.
  2. 2

    Present results and pricing

    Opsio shows per-defect performance with example images, the recommended camera, lens and lighting bill of materials, and how imaging and inference fit your takt. Defects that need another modality, such as 3D for undercut, are called out. You receive a fixed quote for the production station and the managed service.

    You provide
    Quality and production engineers in the review.
    Opsio delivers
    Results report, bill of materials, cycle-time fit, fixed quote.
  3. 3

    Decision and PO/MSA signing

    You decide whether to go ahead. The purchase order and master services agreement are signed, covering station scope, acceptance criteria per defect class, ownership of images and data, and support. Acceptance criteria are written to drop straight into your control plan and PPAP submission.

    You provide
    Purchase order, agreed acceptance criteria, a named station owner.
    Opsio delivers
    Signed MSA and acceptance test plan.
  4. 4

    Installation in the Opsio AI Lab

    Opsio builds and commissions the production station in the AI Lab on the real cameras, lighting, enclosure and edge hardware, and tests it against the acceptance criteria with your parts. The PLC interface is tested against a simulated PROFINET or EtherNet/IP handshake, and PPAP evidence is prepared.

    You provide
    PLC tag list and MES interface details.
    Opsio delivers
    Pre-commissioned station, acceptance results, validation evidence.
  5. 5

    Installation at the customer location

    The pre-commissioned station is installed during a planned line stop, connected to the station PLC, MES and Andon, and run in shadow mode next to your current inspection. Once results meet the acceptance criteria, live rejection is switched on, operators are trained and the active-learning loop starts.

    You provide
    Line access in a planned stop, operators for training.
    Opsio delivers
    Installed station, shadow-mode report, training.

Pricing for the feasibility study, production stations and the managed service is on the AI Visual Inspection service page.

AI visual inspection for Automotive: questions

What are examples of machine vision in the automotive industry?

The main examples are paint finish inspection, weld and seam checks, casting and machined-part inspection, and final-assembly verification. Paint shops use deflectometry tunnels to find orange peel, runs and inclusions. Body shops profile weld beads, hems and sealer beads with laser lines. Powertrain plants check castings for cracks, porosity and burrs. Trim and final lines confirm that clips, connectors, labels and the right part variant are fitted for each VIN.

What is a sagging defect in paint?

A sag is paint that has flowed downward on a vertical surface before curing, leaving a thicker curtain or droplet-shaped edge. It comes from too much film build, low viscosity, slow solvent flash-off or a gun held too close. On a car body, sags and runs show as distortion in reflected light, so a deflectometry tunnel detects them and records their position for spot repair.

What causes orange peel paint on car bodies?

Orange peel comes from paint that cannot level before it cures. Typical causes are poor atomisation at the bell or gun, high viscosity, wrong film build, fast solvent evaporation and booth temperature or airflow outside specification. Because the texture builds up slowly, inspection that measures it on every body lets your paint engineers see a booth or robot drifting before the bodies need polishing.

What are the causes of undercut in welding?

Undercut is usually caused by too much current or voltage, too high a travel speed, or the wrong torch angle, which melts a groove into the base metal at the weld toe that the filler does not fill. In robot cells, worn contact tips and fixture shifts also change torch position. A laser-line profiler measures undercut depth on every bead, so a drifting program shows up as a trend.

Why does porosity occur in casting?

Porosity in castings comes from gas trapped during filling or from shrinkage when the metal cools unevenly. In high-pressure aluminium die casting, air entrapment, die lubricant and hydrogen in the melt cause gas pores, while thick sections cause shrinkage voids. Pores open up when the part is machined, which is why inspection after machining with low-angle light finds them on sealing faces.

How does a paint inspection light tunnel work?

A light tunnel surrounds the moving body with LED panels or bars that create striped reflections on the paint. On a perfect surface the stripes are smooth; orange peel, runs, inclusions and craters bend them. In an automated tunnel, fixed or robot-mounted cameras capture the reflections and software calculates surface slope, so each defect is found, sized and located on the body within the takt.

What is the best automotive paint inspection lighting?

The best lighting for paint is a striped or structured pattern reflected off the surface, used for deflectometry. Glossy paint acts as a mirror, so diffuse light alone hides waviness and small inclusions. Add diffuse channels for scratches and thin spots on matte areas, and use a calibrated spectrophotometric reference, not RGB, for colour. Light colour temperature and stripe pitch are tuned to your paint colours.

How do you measure orange peel in paint?

You measure orange peel by its waviness in long-wave and short-wave ranges, usually with a contact wave-scan instrument on sample spots. Inline deflectometry measures the same surface slope from reflected stripe patterns across the whole body, without touching it. AI models then classify zones as within or outside your appearance targets and trend the values per booth, robot and colour.

How much porosity is acceptable in a weld?

Acceptable porosity is set by the weld standard and your drawing, not by the inspection system. Automotive specifications typically limit pore size, number per length of weld and position, with tighter limits on safety-relevant joints. The AI model is trained on your boundary samples and your acceptance rules, and its decisions on borderline pores are reviewed by your weld engineers during the feasibility study.

What are body in white inspection requirements?

Body in white inspection checks that the welded shell matches its specification before it goes to paint. Typical requirements cover weld presence, position and quality, sealer and hem continuity, gaps and flush between panels, hole presence and surface condition. Each result must be traceable to the body. A vision system reports these per body through MES, so the paint shop receives only bodies that passed.

What is VT weld inspection?

VT weld inspection is visual testing, the most basic non-destructive method, in which an inspector examines the weld surface for porosity, undercut, spatter, cracks and profile faults. AI vision automates the surface part of VT on every weld at line speed. It does not see inside the joint, so internal fusion and penetration still need ultrasonic or destructive testing on your sampling plan.

Is there a visual inspection standard for automotive parts?

There is no single visual inspection standard for all automotive parts; criteria come from OEM customer-specific requirements, drawings, weld and paint specifications, and boundary samples. IATF 16949 requires that inspection methods are defined in the control plan and that their capability is shown, for example by attribute agreement analysis. An AI station is set up against those criteria and documented for PPAP.

Which ISO certification applies to the automotive industry?

The quality management standard for automotive production is IATF 16949, which builds on ISO 9001 with automotive-specific requirements. It covers traceability, control plans, measurement system analysis and PPAP. Opsio does not certify your plant; the inspection system supports your IATF 16949 obligations by storing each result, image and model version against the VIN or part serial.

Book a 20-minute walkthrough for automotive

Pick the defects that cost you most, and we will show you live how a vision model finds them on a line like yours.

  • Live on your defect type, no slides
  • A reply within one business day
  • No obligation
Your call is with
  • Vaishnavi ShreeDirector & MLOps Lead
  • Praveena ShenoyCountry Manager, India

We use your details to answer your request, as described in our Privacy Policy.