Surface Defect Detection on Reflective Parts: Which Lighting Setups Actually Work
From what I've seen, lighting failures on reflective parts usually trace back to one decision made in the first week of a project: choosing whatever light was already sitting in the parts bin, without checking it against the part's actual surface. Chrome, polished aluminium and coated steel all behave differently under the same fixture, and a setup tuned for one will miss defects on another. Reliable surface defect detection on these materials starts with the light, well before the algorithm that processes the image afterwards.
Diffuse Light Beats Directional Light on Curved Metal
Directional lighting throws hard reflections across curved or highly polished surfaces, and those reflections read as noise to most vision algorithms. A dome light or a diffuse ring light spreads illumination evenly and reduces the glare that hides real defects underneath it. It costs more upfront than a basic LED bar and pays for itself the first month a hot spot stops masking a scratch. Plants that switch from a directional bar light to a dome fixture on chrome parts often see their false reject rate drop within the first week of running it.
Dark Field Lighting for Scratches, Bright Field for Contamination
A scratch on a reflective surface shows up best under dark field lighting, where light grazes the surface at a low angle, and the scratch scatters light toward the camera against an otherwise dark background. Contamination, pitting and colour variation usually show up better under bright field lighting, aimed more directly at the surface. Effective surface defect detection setups on reflective parts tend to use both, switching between them by inspection station so a single fixture never has to do both jobs.
Polarising Filters Solve What Angle Alone Cannot
When a part has a mix of matte and glossy regions, angle changes alone will not clear every reflection. A polarising filter on both the light source and the lens cuts specular glare without needing to chase the light around the part. It is a cheap add-on that gets skipped constantly because it is not on anyone's initial parts list. A crossed-polariser setup, one filter on the light and one on the lens rotated ninety degrees against it, removes almost all specular glare on a smooth metal surface.
Where the Algorithm Comes In, After the Light Is Right
No amount of clever software fixes a badly lit image. Good ai based defect detection models still need consistent, well-separated pixel values between a real defect and a reflection artefact to work reliably. Vendors who lead with the model before asking about your lighting setup are solving the wrong half of the problem first. An AI-based defect detection model trained on well-lit sample images performs far worse on your floor once the production lighting stops matching those training images.
Colour Temperature Consistency Matters More Than Brightness
A light source that shifts colour temperature as it warms up, common with cheaper LEDs run continuously for a full shift, can slowly drift a system's baseline without anyone noticing until rejects start climbing. Buy lights rated for colour stability over a full duty cycle, and check the manufacturer's stated drift figures as closely as you check brightness before deciding.
Document What Each Fixture Actually Fixes
Keep a simple log of which lighting configuration solved which defect type, with a photo of the setup and a sample image beside it. Six months later, when a new part variant arrives with a similar surface, that log saves a full week of re-testing lighting from scratch. Plants that skip this documentation tend to relearn the same lighting lessons every time a new product line ships, because the knowledge lived only in one engineer's memory and left when that engineer moved on.
Test the Lighting Before You Test the Vendor
Before scoring any vendor's software, run your own reflective parts under a few lighting configurations and photograph the results. If a defect is invisible or ambiguous to your own eye in the photo, no model downstream will reliably catch it either. Fixing the light is the cheapest change in the whole project, and most teams try it last, after everything else has already failed. Move it to the top of the list on the next reflective-parts project and the software evaluation that follows gets noticeably easier.