Surface defect inspection uses carefully designed lighting and cameras to reveal scratches, dents, cracks, pores, stains and coating faults, then classifies them with rule-based tools or deep learning. It gives every part the same careful look, at production speed.
Surface defect inspection checks the outside of a part for flaws that affect its function or appearance. It is one of the hardest jobs for people to do consistently, because many defects only show at a particular angle of light.
Machine vision solves that with lighting engineered to make the defect visible, and software that decides whether what it sees is a defect or acceptable variation. Deep learning is especially useful here, because it can learn what a good surface looks like and flag anything unusual.
Parts pass on a conveyor, rotate in front of the camera, or are moved by a robot so every face is seen.
The lighting geometry is chosen to make each defect type stand out: low-angle light for scratches, diffuse dome light for shiny parts, backlight for cracks in transparent material.
Rule-based tools find predictable defects; deep learning models flag anything that differs from good parts.
Defects are classified and sized, and the part is accepted, rejected or sent for review against your criteria.
Defect counts by type and position show which upstream process needs attention.
Light grazing the surface makes scratches, dents and embossing throw bright or dark edges, even on flat parts.
Even, shadow-free light removes glare on shiny, curved or metallic parts so real defects are not confused with reflections.
Several images lit from different directions are combined to reveal surface shape, making shallow dents and bumps visible.
A model trained mainly on good parts learns normal variation and flags anything unusual, which suits defects that are rare or unpredictable.
Two example project scopes showing how the station, the checks and the outputs are defined. Every plant is different, so the final configuration is confirmed after a feasibility study on your own parts.
The situation: A rubber products exporter relies on visual inspection and receives cosmetic complaints from overseas buyers.
The situation: A metal fabricator ships painted parts and needs consistent cosmetic grading before packing.
Industrial cameras, lenses, line-scan cameras and code readers are available through our machine vision catalog, quoted with delivery to Sri Lanka. We select the hardware for your application during the feasibility study.
Browse the hardware catalogEvery product, surface and line behaves differently under a camera, so every project starts with evidence from your own material.
A site visit to understand your product, line speed, quality criteria and where the check should sit.
We image your good and defective samples and show what the system can reliably detect. This is the go or no-go.
One station on one line, running beside your inspectors until results match and your team trusts them.
More lines, integration with your PLC and systems, and local support from our Sri Lankan engineering team.
Non-contact measurement of lengths, diameters, hole positions, gaps and profiles with telecentric optics and 3D laser profiling. Measure every part in-line, pass or fail it against tolerance, and log data for SPC.
Camera-based fabric inspection for woven and knit rolls. Line-scan cameras detect holes, stains, slubs and dropped stitches across the full width, map every defect and grade rolls to the 4-point system.
Sort fabric rolls, cut panels and washed garments into buyer shade bands with CIELAB measurement and 555 codes.
Every machine vision application we design and build, in one place.
Yes, with the right lighting. Glare is the main problem on shiny surfaces, so diffuse dome lighting or photometric stereo is used to suppress reflections and make scratches and dents stand out.
Rule-based tools suit defects that are predictable in size, shape and contrast. Deep learning suits variable, cosmetic or rare defects on textured or natural surfaces, where writing rules for every case is impractical.
For a feasibility study, a set of good parts and as many examples of each defect type as you can collect. Deep learning anomaly detection can start mainly from good parts, which helps when defects are rare.
Yes, by using several cameras, rotating the part, or having a robot present each face to the camera. The right approach depends on part size, shape and line speed.
Borderline parts can be routed to a review station where an inspector sees the image and makes the call. Those decisions can then be used to refine the model.
Tell us what you need to check and share some good and faulty samples. We will show you what a camera can reliably see before you commit to anything.
Photographs show representative plants and equipment, not Cerox installations or client sites. Diagrams by Cerox Engineering. Photo credits, via Wikimedia Commons: Metal sheet in production by Shixart1985, CC BY 2.0; Blowhole defect in a casting by Orion Lawlor, Public domain.