Feasibility guide

Machine vision for difficult products

Reflective, transparent, very dark, deformable and naturally variable products are not one problem. Separate the imaging challenge from the inspection or guidance decision and test the worst credible conditions.

Discuss your application Read the guide

Hands using digital calipers on a metal component.

Manual inspection shown: representative trials establish whether the required feature can be automated.

SurfaceGloss, transparency, darkness and low contrast

PresentationOverlap, movement, deformation and changing background

EvidenceRepresentative samples under real production conditions

Direct answer

Can machine vision inspect reflective, transparent or black products?

Often, but the imaging method must reveal the required feature consistently. Lighting angle, wavelength, polarisation, background, optics, exposure and sometimes 3D or learned analysis are evaluated with representative samples. Feasibility should be proven rather than assumed.

01 / Surface and material

Create contrast before choosing the algorithm

A shiny surface can reflect the light source instead of revealing a feature; a clear product can show the background; a dark surface can return very little light. The solution may rely on geometry, diffusion, polarisation, backlighting or a different sensing method.

Diffuse and dome lighting for specular surfaces
Backlight for silhouette, edge and presence
Cross-polarisation to reduce selected glare
Bright-field or dark-field geometry for surface features
Wavelength and filter trials where the material response differs

02 / Real variation

Test the difficult production range, not one ideal part

Product finish, moisture, dust, temperature, supplier, batch, movement and ambient light can change the image. The trial set should include good, defective and borderline examples across those conditions.

Multiple batches, colours and finishes
Fastest speed and widest presentation range
Contamination, condensation or normal process residue
Worst-case overlap, deformation and orientation
Uncertain-result and operator-review policy

Side-by-side comparison

Match the method to the production constraint.

Use the comparison as a screening tool; the final architecture follows from representative products, timing and acceptance evidence.

Typical starting points only. The required feature and representative trial determine the final method.
Challenge	Why it is difficult	Possible engineering routes	Evidence to provide
Reflective metal or film	Glare can hide edges, print or surface defects	Diffuse illumination, dark-field, polarisation, angle control	Finish range, scratches, curvature and production lighting
Transparent product	Background and internal reflections dominate the image	Backlight, controlled background, edge methods, specialised geometry	Material, thickness, contents, bubbles and required defect
Black or absorbent surface	Low returned signal and weak contrast	Higher-efficiency lighting, wavelength trials, silhouette or 3D	Surface texture, finish, temperature and smallest feature
Deformable or natural product	Shape and appearance vary within acceptable product	Robust geometric features, AI where justified, explicit tolerance classes	Representative good range, defects and borderline decisions
Overlapping or random parts	Features and grasp points may be occluded	3D scene capture, alternate views, presentation control, recovery	Container, fill level, part interaction and required pick rate

Decision pathway

Turn uncertainty into a test plan.

A useful feasibility trial reproduces the image, timing and decision conditions that the production system will face.

Define

State the smallest feature and the required production response.

Collect

Provide the full good, defect and difficult variation range.

Trial

Test lighting, optics, presentation and processing at realistic speed.

Decide

Record proven limits, uncertainty handling and the next engineering step.

Questions manufacturers ask

Questions about difficult surfaces and variation.

Use these answers to frame the first application review and the evidence needed next.

Will a higher-resolution camera solve poor contrast?

Not by itself. More pixels cannot recover a feature that the lighting and optics do not make visible. Image geometry and contrast should be solved before resolution is increased.

Can AI remove the need for controlled lighting?

No. AI can help with complex visual variation, but lighting changes still alter the input data. Stable image acquisition remains important for validation and long-term performance.

How should difficult products be sampled?

Include different batches, suppliers, colours, finishes, speeds, orientations, contamination states, known defects and borderline cases—not only clean ideal products.

What happens if the system is uncertain?

The machine can route uncertainty to reject, review, alarm or line stop according to the risk. The response should be defined before acceptance testing.

Connected guidance

Plan the imaging trial.

Explore the technologies and buyer guidance most relevant to this application.

Start with the production evidence

Show us the production task and required outcome.

Use the project brief form, or email representative product images or video with the target output and required decision. We will define the next technical step.

Send your project brief 01865 416830 sales@oxfordindustrialautomation.co.uk