Use Cases

Surface Defect Detection

AI surface defect detection for metal parts: finds scratches, dents, cracks, burrs and coating faults, learns from good parts and rejects via your PLC.

Surface defect detection finds scratches, dents, cracks, burrs, pits and coating faults on metal and other manufactured parts before they leave the line. DeepVision images every part, compares it with what a good part looks like, marks where the problem is, and sends a pass or fail decision to your PLC so the reject happens automatically.

For background, read our guides to defect detection with AI and machine learning and AI visual inspection in steel and metal manufacturing.

100%accuracy on a Tier-1 supplier’s body-in-white line (hole, bolt and perimeter checks)
≥98%detection on a 10-camera flat-panel board surface inspection line

On the body-in-white line, DeepVision counted holes and bolts on each sheet, and a side camera watched the full sheet perimeter for edge defects, cracks and deformations. Read the sheet metal inspection case study and the board inspection case study, which covers boards up to 1240 × 2445 mm with 100% board coverage.

What surface defect inspection checks

DefectWhat it looks like on the part
Scratches and scuffsFine lines from handling, fixtures, conveyors or die contact. Often only visible at certain light angles.
Dents and dingsLocal depressions or bumps that change how the surface reflects light, even when the paint or finish is intact.
Cracks and splitsHairline cracks at bends, draws, holes and edges, where the material has been stretched hardest.
Burrs and edge defectsRaised material, rough or torn edges after cutting, punching or trimming, and chipped or deformed edges.
Pits and porositySmall holes or clusters of voids in castings, welds, coatings or rolled surfaces.
Coating and paint defectsRuns, sags, bare patches, orange peel, blisters, inclusions and uneven coverage.
Stains and contaminationOil marks, rust spots, discolouration, residue and foreign material left on the surface.
Wrinkles and deformationWavy or buckled areas on pressed panels and sheet parts.

Why manual and sample checks miss surface defects

A trained inspector finds most surface defects under good light on the parts they actually look at. But hairline cracks and light scratches on bright metal are easy to miss late in a shift, and on large parts a rushed glance covers the obvious faces but not the edges. Sample checks only see the parts that are pulled, so a worn die, a damaged fixture or a drifting paint gun can produce a run of faulty parts between two samples.

Judgement also varies between inspectors, and manual checks leave no image behind for when a customer complains.

Reflective and textured surfaces: why lighting matters

Polished and bright metal acts like a mirror: the camera sees glare, a scratch can vanish, and a harmless reflection can look like a defect. Brushed, blasted, cast and textured surfaces have the opposite problem, where a real defect has to stand out against a busy background pattern.

That is why the imaging set-up matters as much as the AI model. Depending on the part, the station may use diffuse lighting to suppress glare, low-angle lighting to make scratches and raised edges cast a shadow, or structured light so dents and waviness show up as a break in a reflected pattern. Camera, lens and lighting are fixed at the station so every part is imaged the same way, which is what makes small deviations stand out.

How DeepVision detects surface defects

  1. Capture. One or more cameras under controlled lighting image every part as it passes, covering the faces and edges that matter. Large parts can be covered by several cameras whose images are stitched together.
  2. Learn from good parts. DeepVision uses anomaly detection. It is trained on as few as 200 images of good parts, with no defect labelling, and learns what normal surface looks like. Rare defects you have never photographed can still be flagged because they don’t match. Read more about anomaly detection without defect images.
  3. Locate. For each part, the model produces a heat map showing where the surface differs from normal. Operators and quality engineers see exactly where the scratch, dent or crack is, not just a fail flag.
  4. Decide. Inference takes about 22 ms on edge hardware next to the line, and each part gets a pass or fail decision against thresholds you set.
  5. Reject. The decision goes to your Siemens, Allen-Bradley or Mitsubishi PLC, which drives the existing reject gate, diverter or line stop.
  6. Record. Every part’s image and result is stored, so a reject, a customer query or a root cause investigation can be traced back to the actual part.
No defect library neededTraining starts from good parts, so you don’t have to collect and label hundreds of scratched or cracked samples first.
See where, not just whetherThe heat map points to the defect area, which speeds up re-inspection and helps trace the cause back to a die, fixture or process step.
Runs on siteInspection runs on edge hardware in the plant, with no cloud connection. See security and ISO/IEC 27001:2022.

Where surface defect detection is used

  • Automotive stamping and body-in-white: pressed panels, brackets and sheet parts checked for cracks, edge defects, wrinkles and missing features. See AI inspection for automotive.
  • Machined and cast components: scratches, burrs, pits and porosity on finished faces and edges.
  • Painted and coated parts: coverage, runs, blisters and contamination after paint, powder coat or plating.
  • Flat panels and boards: large-format surfaces imaged by a camera array for full coverage.
  • Steel and metal processing: sheet, coil and fabricated products.

Frequently asked questions

Can AI detect scratches on shiny metal?

Yes, provided the lighting is designed for the surface. On bright or polished metal, diffuse or low-angle lighting is used so a scratch shows up as a consistent change in the image rather than being hidden by glare. We assess lighting with your parts before quoting accuracy.

Do we need images of every defect type to train it?

No. DeepVision learns from images of good parts, so defect samples are not needed to get started. Defect samples are useful for validating the system before it goes live.

Can it tell a harmless mark from a real defect?

Whether a small mark is a reject depends on the part and where the mark sits. The reject rule is agreed with your quality team during setup and checked during the parallel run, and the heat map shows exactly what was flagged. See our accuracy methodology for how we measure results.

Scratches, cracks or dents reaching your customers? Read the sheet metal case study, see DeepVision AI visual inspection, or book a demo with your own parts.

Scroll to Top