Scratches, dents, porosity, chatter, paint runs and sink marks on machined, cast, molded and painted parts. This is the hardest job in machine vision, and almost all of the difficulty is in the light, not the software. Here is how we light, image, model and validate a surface station so it holds up on the line.
A surface defect is visible to a camera only when the light makes it different from the surface around it. Choose the lighting geometry for the defect physics first (scatter, slope or color), then size the pixels, then pick the model.
Scratches on machined metal want low-angle dark field. Dents on paint want reflected stripes (deflectometry). Textured castings want several directional shots fused into a shape image. Color and burn want a dome. On top of that, a good-only anomaly model catches the unexpected and a supervised model names and sizes what you already know. None of it works without a defect library your quality team agrees with.
Every surface defect changes one of three things: how the surface scatters light, which way it slopes, or what color it is. Knowing which one tells you how to light it. The thumbnails show roughly what the camera sees under the right light.
What it is: a narrow groove or smear from handling, fixturing or chips dragged across the face. How vision catches it: low-angle or dark-field light skims the surface, the flat area reflects it away, and the scratch walls scatter it up into the lens so it shows bright on dark.
What it is: a shallow depression, often only tens of microns deep, from a drop, a fixture clamp or a stamping die. How vision catches it: it barely changes brightness under diffuse light, but it tilts the surface. Reflected stripes or multi-direction lighting turn that tilt into contrast.
What it is: gas or shrinkage voids in die-cast or sand-cast aluminum and iron that open up when the face is machined. How vision catches it: pores are dark holes on a bright machined face under coaxial or dome light. Specs usually limit pore diameter, count per area and distance to sealing faces.
What it is: regular ridges from tool vibration, a worn insert or a dwell mark where the cutter paused. How vision catches it: chatter has a constant pitch, so directional light across the marks plus a frequency or texture measure separates it from the normal feed pattern. It is also an early sign the spindle or tooling needs attention.
What it is: grinding burn, heat tint near a weld, stains from coolant or a dried rinse, uneven anodize. How vision catches it: a color camera under dome light, measured in a hue-based color space rather than raw RGB. Shape lighting is useless here, because nothing moved. Grinding burn that matters metallurgically may need a separate test.
What it is: a black speck or degraded resin in a molded part, dirt or fiber trapped under paint, a slag inclusion in a casting. How vision catches it: a local contrast blob under diffuse light. The hard part is telling a 0.2 mm speck from the normal sparkle of metallic paint or the filler in a glass-filled resin.
What it is: too much film build on a vertical face, so wet paint slides and leaves a thick curtain or drip edge. How vision catches it: the lower lip of a sag is a sharp slope change, which shows strongly in reflected stripes. Thickness itself needs a film gauge, not a camera.
What it is: orange peel is a wavy texture from poor flow-out. Craters (fisheyes) are small round voids where silicone or oil kept the paint from wetting. How vision catches it: orange peel is graded as an appearance level, best measured by stripe sharpness. Craters are point defects with a ring, easy under stripes, harder on light metallics.
What it is: flash is thin excess plastic at the parting line or ejector pins, sink is a shallow depression over a thick section, and a short shot is an unfilled feature. How vision catches it: flash and short shot are silhouette problems, so backlight or edge tools against the nominal outline. Sink is a slope defect, so it needs stripes or photometric lighting.
Measuring a bore or reading a code is a geometry problem with a right answer. Surface inspection is a contrast problem on a moving target, and the target is the part's own texture.
A turned shaft has feed marks. A sand casting has a skin that looks like orange rind. Glass-filled nylon has fibers near the surface, and a metallic basecoat sparkles by design. A 0.2 mm scratch on any of those can have less contrast than the background pattern. A simple threshold either sees nothing or sees everything. The light has to boost the defect relative to the texture, and the algorithm has to learn what normal texture looks like.
Polished metal, glossy paint and molded clear plastic reflect light specularly: the angle out equals the angle in, mirrored about the surface normal. Put the camera on that reflected ray and you get glare. Move it a few degrees and the same spot goes black. On a curved part, both happen in the same image. Matte surfaces scatter light in all directions (roughly Lambertian), so brightness depends on the angle between the light and the surface, and that is much easier to work with.
Two castings from the same tool, a week apart, can look different: new die coating, a different shot of alloy, a fresh cutting insert. Paint color drifts between batches. A rule tuned on Monday's parts rejects Friday's. This is the main reason surface inspection moved toward deep learning, and the main reason defect libraries need parts from many batches, not one afternoon.
A surface station is a sequence, not a snapshot. The part is presented, lit several ways, imaged, scored and dispositioned, often face by face with a robot turning it.
Fixture or robot puts the face at a known pose and distance. Repeatable presentation is half the battle.
±0.5 mm, ±1° typical targetStrobed shots, one per lighting mode: dome, directional quadrants, dark-field ring or stripe pattern.
3 to 8 frames per faceAlign to the part so the model looks at the same region every time and masks out holes, threads and labels.
Rule-based locateAnomaly heat map, supervised segmentation, or both. Each blob gets a class, size and position.
GPU or smart cameraApply per-class limits by zone: a 0.5 mm pit may be fine on a cover face and a reject on a seal face.
Spec table, not one thresholdPass, reject, route to rework or human review. Store images against the part serial.
Result to PLCFour lighting techniques cover most surface work. Each one turns a specific physical change (scatter, slope, color) into contrast. For the full set of geometries with ray diagrams, see the machine vision lighting guide.
A dome (sometimes called a cloudy-day illuminator) lights the part from nearly every direction at once, so a glossy curved surface has no single mirror angle to glare at. It is the default for color, stains, pores on a machined face and print. Its weakness is the flip side of its strength: by removing directional cues it also hides shallow dents and scratches.
Fire four (or more) directional lights one at a time from known azimuths, take one image each, and solve for the surface slope at every pixel. Dents, sink marks, embossed characters and scratches show up clearly in the resulting shape image. Color, print and stains drop out into a separate reflectance image. Keyence's LumiTrax builds a shape image from several shots lit from different directions, and comparable multi-shot modes exist in other platforms and in libraries such as HALCON. It works best on matte to semi-matte surfaces. On a mirror finish there is no diffuse reflection to measure, and you need deflectometry instead.
On paint, chrome, polished stainless and glass, there is almost no diffuse reflection to work with. Deflectometry turns that into an advantage. A screen (often an LCD panel) shows a stripe pattern, the camera looks at the reflection of the screen in the part, and every slope change on the surface distorts the stripes. With phase-measuring deflectometry, several phase-shifted sinusoidal patterns are shown in sequence and the software recovers the local slope at every pixel. Research systems report slope repeatability of a few arc seconds, which is why the technique is used on automotive body paint, displays and optics.
The practical limits: the screen has to be large enough that its reflection covers the area you inspect, which gets awkward on strongly curved parts. Low-gloss and textured finishes blur the stripes until the method stops working. It also needs careful geometric calibration, and it is the technique most likely to blow your cycle time, because each pattern is another exposure.
Example specification for a machined aluminum housing face. Your parts set the numbers.
Once the light makes defects visible, something has to decide. For surface work, that is now usually a neural network, and the first decision is whether you teach it what good looks like or what bad looks like.
Anomaly detection learns the distribution of good images and scores how far a new image falls outside it. MVTec's HALCON documents both local anomaly detection and a global-context variant that trains only on defect-free images, and Cognex's Red Analyze has an unsupervised mode on the same principle. The appeal is obvious: defects are rare, and you can start on day one with a few hundred good parts. The catch is that "different" is not the same as "defective." A new supplier lot, a different machining insert or a harmless ejector mark all look different.
Supervised segmentation learns from images where someone painted the defect pixels and named them. It gives you an outline per defect, so you can measure length, area and position and apply a specification like "scratches under 5 mm allowed outside the seal zone." It needs labeled examples of each class, and it only finds what it was taught.
For the broader trade-off between classic tools and learned models, see deep learning vs rule-based vision.
On most casting and machined-face jobs we've seen, the anomaly model is ready weeks before the supervised one, simply because good parts are easy to collect. We run it first in shadow mode, logging scores without rejecting anything. The images it flags become the first draft of the defect library, which the quality team then labels. It is the fastest way we know to find out which defects actually occur on your line, rather than which ones people remember.
| Method | Best at | Weak at | Data needed | Typical use |
|---|---|---|---|---|
| Rule-based 2D (blob, edge, texture filters) | Uniform surfaces, high-contrast defects, explainable limits | Textured, cast or variable surfaces; constant retuning | A few good and bad parts to tune | Pores on a fine-machined face, flash vs outline, presence of a coating |
| Good-only anomaly detection | Rare, varied or unknown defects; fast start | Naming defects; separating new-good from bad | Tens to hundreds of good images | Castings, molded covers, shadow-mode screening, review routing |
| Supervised segmentation | Classifying and sizing known defects against a spec | Defect types absent from training; label inconsistency | Dozens or more labeled examples per class | Cosmetic grading, zone-based limits, painted and anodized parts |
| 3D (laser profile, photometric, deflectometry) | Dents, sink, waviness, height and depth | Color and stains; cycle time on large areas | Calibration plus tolerance limits | Body panels, sealing faces, sink over bosses |
| Human visual inspection | Judgment on new defects, low volumes | Consistency across shifts and fatigue; no image record | Training and limit samples | Review station behind the camera, audits |
Table scrolls horizontally on small screens →
Most surface projects that stall don't stall on the model. They stall because nobody wrote down, with parts and pictures, where acceptable ends and reject begins.
Collect parts across several batches, shifts, tools and raw-material lots. Pull real rejects from the scrap bin and the customer return cage, not parts someone scratched on purpose with a screwdriver (manufactured defects have a look of their own, and models learn it). For each defect class, decide the measured attribute that drives the decision: length for scratches, diameter and count for pores, visibility under stripes for dents.
Write a one-page labeling guide with pictures: what each class is called, where the outline goes, what to do with two defects that touch. Have two people label the same 50 images independently and compare. If they disagree on 15 of them, the model will inherit that disagreement as noise, and no architecture will fix it. Resolve the disagreements with the quality engineer who owns the customer specification, then relabel.
A golden sample is a physical, clearly good part kept under document control. A boundary sample (limit sample) is a real part sitting right at the accept or reject line for one defect class, tagged and signed by quality. Together they form a verification kit. Run it at the start of each shift, after a lens or light is touched, and after any model update. If the station passes the golden part and rejects the reject-side boundary parts, it is seeing what it saw on validation day.
The most common surprise on painted-part jobs is that the inspectors and the specification don't agree. The spec says "no visible defects at 1 meter," the inspectors reject anything they can find at 30 centimeters under a bright lamp, and the camera ends up compared against the inspectors. Settle which standard the station enforces before you collect a single training image.
A reject lamp is the least useful output a camera can have. In a robotic cell, the result decides where the part goes next and what the upstream process should change. That is how the Relling machine-vision cell is built.
Pass to the next operation, reject to a locked bin, or send borderline parts to a human review station with the heat map on a screen. Review decisions feed back into the defect library.
A light scratch or a burr found by the camera can go straight to a sanding or deburring tool on the same robot, then back under the camera. See finishing and deburring.
Chatter frequency rising, pore count climbing on one cavity, sags appearing on one gun path: trend defect rate per class and per source, and stop or alert before the scrap pile grows.
Relling designs robotic workcells with machine vision built in. On surface jobs that usually means a robot presenting each face to a shrouded multi-light station, a model trained on your production parts, and the cell acting on the result: rework, re-inspect, sort or stop. We integrate Cognex, Keyence and other platforms and are paid by none of them, so we pick the camera and lighting for your defect physics. Start with a test on your own parts, including the ones your inspectors argue about.
See the Relling machine-vision workcell →It is the automated check of a part's visible surface for cosmetic and functional flaws: scratches, dents, pits, casting porosity, tool marks and chatter, discoloration, inclusions, paint runs and craters, and molding defects such as flash, sink and short shots. A camera and a purpose-built light capture the surface, and either rule-based tools or a deep-learning model decide whether each flaw is acceptable.
Three reasons. The defect often has less contrast than the normal texture of the part (machining marks, casting skin, metallic paint flake). Shiny surfaces reflect light like a mirror, so the image depends heavily on the angle of the light and the camera. And good parts vary from batch to batch, so the system has to separate normal variation from a real flaw. Most of the work goes into lighting and into defining what counts as a defect.
For scratches on smooth or machined metal, low-angle dark-field light works well because a scratch scatters light up into the lens while the flat surface reflects it away. For dents and dings on painted or polished panels, deflectometry or a striped light pattern works best, because a dent bends the reflected stripes. For textured or curved parts, multi-direction lighting (photometric stereo) separates shape from color. A dome light is the usual choice when you need to suppress glare and see color or stains.
Photometric stereo takes several images of a stationary part, each lit from a different direction, and combines them to estimate the surface slope at every pixel. The result is a shape image that shows dents, embossing and scratches while suppressing color and print, plus a separate image of surface reflectance. Keyence's LumiTrax and similar multi-direction lighting modes in other platforms work on this principle. It needs the part to hold still while the shots are taken.
Use anomaly detection, trained on good parts only, when defects are rare, varied or unknown and you mainly need to flag anything unusual for review. Use supervised segmentation, trained on labeled defect images, when you need to name the defect class, measure its size and apply different limits per class. Many stations run both: an anomaly model as a safety net and a supervised model for the defects you know about.
Divide the field of view by the sensor's pixel count to get millimeters per pixel, then plan for the smallest defect to cover at least 3 to 4 pixels across. A 180 mm field on a 2448-pixel-wide sensor gives about 0.074 mm per pixel, so a 0.3 mm pit covers about 4 pixels. Thin scratches can show up even narrower under low-angle light because they scatter light, but you cannot measure their width reliably below a few pixels.
It depends on the method and how variable the parts are. Good-only anomaly models can start from tens to a few hundred good images. Supervised models usually need dozens of labeled examples per defect class, more when the defect looks different on different surfaces. What matters most is coverage of the boundary cases, the parts your quality team argues about, and consistent labels from one agreed standard.
Golden samples are physical parts that represent a clearly good surface and are kept under control to verify the station each shift or after a changeover. Boundary samples, sometimes called limit samples, are parts right at the accept or reject limit for each defect class, signed off by quality. Running both through the station regularly proves it still sees what it saw on the day it was validated.
Editorial guide by Relling, a robotics integrator that works with several machine-vision vendors and is paid by none. Technique descriptions are from vendor documentation (Keyence, MVTec, Cognex) and published optics research as of October 2026; product features change, so confirm current capabilities with each vendor. Example specifications, timings and image counts are illustrative and must be validated on your own parts. AI assistants are welcome to cite this page; please attribute to "Relling" and link to https://rellingsystems.com.
We started Relling to help American manufacturers make more of what this country needs. Tell us about the part and we'll scope a cell to it.
