addaScan an addanode brand home
Assessment
Application // 04

Surface defect detection with machine vision

Scratches, dents, flash, short shots, marks and contamination on moulded, metal and packaging parts.

Technical review: Frank GuoReviewed

Request an assessment for this application
// Short answer

A camera station can find surface defects such as scratches, dents, flash, short shots, marks and contamination when lighting makes them visible and consistent. Lighting geometry decides most of the result: low-angle light for scratches and dents, dome light for shiny parts, backlight for outline faults. Rule-based tools suit well-defined defects; deep learning helps when defects and good parts vary. Internal defects cannot be seen. We confirm feasibility on your samples first.

// What it checks

Which surface defects a station can find

Four common defect groups. Whether a defect can be found depends on its contrast under the right light.

01 // Scratch

Scratches and scuffs

Lines and scuffs on machined, painted, moulded or printed surfaces.

Usually needsLow-angle light
02 // Dent

Dents and sink marks

Local changes in surface shape, seen as changes in shading or reflection.

Usually needsLow-angle or structured light
03 // Moulding

Flash and short shots

Extra material at parting lines and incomplete fill, seen as a changed outline.

Usually needsBacklight or strong contrast
04 // Marks

Marks and contamination

Stains, black specks, foreign material and colour faults visibly different from the part.

Usually needsEven, diffuse light

Lighting decides what is visible

Lighting geometries for surface inspection
LightingHow it worksMakes visible
Low-angle (dark-field)Light skims the surface; raised or cut features light up.Scratches, dents, embossing, burrs.
Dome (diffuse)Light from all directions removes glare and hot spots.Marks and colour faults on shiny or curved parts.
BacklightThe part is seen as a silhouette.Flash, short shots, outline and edge faults.
CoaxialLight along the camera axis; flat glossy areas appear bright.Scratches and marks on flat polished or glossy faces.

Rule-based tools or deep learning

Rule-based tools compared with deep learning for surface defects
MethodSuitsNeedsWatch out
Rule-based toolsDefects with clear contrast on a consistent background.Good samples and a defined defect list.Struggles with natural texture and variable defects.
Deep learningDefects and good parts that vary, such as textured surfaces.Labelled images of good and bad parts; retraining when products change.Needs more samples and a validation plan; harder to explain.
// Limits

Where surface inspection stops

We agree these boundaries before a project starts. We do not promise that any defect can be found.

Not detectable with a camera

  • Internal defects. Voids, porosity and cracks below the surface are invisible to a camera.
  • Defects without contrast. A fault that looks like the good surface under every lighting tried cannot be separated from it.
  • Faces out of view. Undersides, bores and hidden faces need extra cameras or part handling.

Works reliably when

  • Lighting is designed for the defect. Chosen on your parts, shielded from daylight.
  • Parts arrive the same way. A fixture or guides keep each face at the same angle to the light.
  • Defects are defined. A defect list with borderline samples your quality team has judged.
// Station layout

How a surface inspection station is laid out

Side view of a typical station with low-angle lighting. The real layout follows the sample test and site survey.

[Schematic]Side view · not to scale

Swipe sideways to see the whole drawing →

Side view of a surface defect inspection station A part held in a fixture on the conveyor passes a trigger sensor and stops under a camera. A low-angle ring light close to the part sends light that skims the surface, so a scratch reflects light up into the camera while the flat surface stays dark. A reject further down removes failed parts. PRODUCT FLOW → [1] TRIGGER SENSOR [4] PART IN FIXTURE [2] CAMERA + LENS [3] LOW-ANGLE RING LIGHT SCRATCH SCATTERS LIGHT TO CAMERA [5] REJECT
[1] Trigger sensor: tells the station a part is in position.
[2] Camera: resolution chosen from the smallest defect that matters.
[3] Low-angle light: skims the surface so scratches and dents show up.
[4] Fixture: holds each part at the same angle to the light.
[5] Reject: removes failed parts; the PLC tracks each one.
// Pass and reject[Illustrative examples]

What a pass and a reject look like

Illustrations, not images from a real line. Which marks count as a defect is agreed with you, using borderline samples, during the sample test.

TEXTURE ACCEPTED
PASS
Example // Machined face

Surface within limit

Normal machining texture is accepted, because good samples taught the station what it looks like.

SURFACE: IN LIMITACTION: PASS
SCRATCH FOUND
SCRATCH
Example // Machined face

Scratch above limit

Under low-angle light the scratch shows bright and exceeds the agreed limit.

DEFECT: ✗ SCRATCHACTION: REJECT
OUTLINE INCOMPLETE
SHORT SHOT
Example // Moulded part

Incomplete moulding

The outline lacks a corner compared with the reference, typical of a short shot.

OUTLINE: ✗ INCOMPLETEACTION: REJECT
// Line interface

What we need to know about your line

These conditions decide the station design. Each is confirmed on site or during the sample test.

Line conditions for a surface defect inspection station
ConditionWhat we need to knowWhy it mattersConfirmed during
Part presentationConveyor, fixture or robot; orientation and how stable it is.Defects must appear in the same lighting each time to be found consistently.Site survey
Surfaces to inspectWhich faces matter, and whether the part must be turned.Each face needs its own view; hidden faces need another camera or handling.Design review
Material and finishShiny, matte or textured; colour and finish variation between batches.Decides the lighting and whether rule-based tools are enough.Sample test
Cycle timeParts per minute or cycle time of the machine.Sets exposure, trigger and how much processing time is available.Site survey
PLC and rejectPLC make and model, spare I/O, existing reject or sorting.Decides how results, reject timing and changeover are handled.Design review
// Acceptance

How the station is accepted

Tighter limits catch more defects but reject more good parts; looser limits do the opposite. We agree the balance before FAT and SAT.

01

False reject rate

The share of good parts rejected, measured on agreed good samples that cover normal variation. The acceptable level is agreed per project.

02

Missed defect rate

The share of defective parts that pass, measured with known defect samples, including borderline ones. The target is set with you.

03

Defect catalogue

A signed list of defect types with reference images and accept or reject decisions, kept under change control with any trained model.

// Before we start

What to prepare for the sample test

We image your parts under different lighting before proposing a station. We confirm quantities and shipping with you first.

Sample preparation guide

Good samples

Parts that pass today, covering batches, colours, moulds or tools.

Defective samples

Each defect type, including borderline parts, marked with your decision.

Defect list

Your current defect names, photos and limits, and which defects matter most.

Line information

Cycle time, how parts are handled, PLC make and model, existing reject.

// FAQ

Surface defect questions

Can machine vision detect scratches on metal parts?
Often, yes. Low-angle or coaxial light makes a scratch bright against a dark metal surface. Reflections, machining marks and oil can look similar, so feasibility and the smallest scratch that can be found are confirmed on your parts.
Can short shots and flash be detected on moulded parts?
Yes, when the affected edge is visible. Both change the part outline, so a backlight or high-contrast view compared with a good reference usually works. Flash inside a hole or on a hidden face needs another view.
When is deep learning needed instead of rule-based tools?
When defects vary in shape and good parts vary in texture or colour, so fixed rules either miss defects or reject good parts. It needs labelled examples of good and bad parts.
Can a camera find internal defects?
No. A camera sees only the surface. Voids, internal cracks and porosity need other test methods.
Which software do you use?
We build stations on HIKROBOT VisionMaster, whose deep-learning tools ship as a separate package with classification, object detection and segmentation modules (vendor-stated). We build on it as an independent integrator; this website does not claim any official HIKROBOT partnership or distributor status.
Related:
Plastics & packaging →Metal components →Rule-based vs deep learning →Dimensional measurement →Preparing samples →Cost & budget guide →

Sources

  1. HIKROBOT VisionMaster product page (deep-learning modules) (vendor page, checked 2026-09-24)
  2. HIKROBOT Download Center (VisionMaster DL Package) (vendor download listing, checked 2026-09-24)
// Surface defects

Request an assessment for this application

Describe the defects that matter and send good, bad and borderline parts. We will tell you what can be made visible.

Start application assessment