Computer Vision + Product

Machine Vision Products Need a Lighting Contract

• Christopher Mitchell •5 min read

Machine vision products need measurable lighting limits, image-quality checks, and recovery rules before model accuracy becomes an operating promise.

The inspection station rejects clean labels every morning and passes them after lunch. The model, camera, and labels have not changed.

Sunlight reaches the conveyor through an open bay door before noon. Later, the overhead fixtures dominate the scene. The classifier is blamed because it owns the final answer, although the input image has become a different product.

Machine vision teams often treat illumination as installation work. Then a shadow, reflective wrapper, dimming LED, or shifted fixture changes contrast enough to break a system that looked finished.

A production vision product needs a lighting contract: measurable limits for illumination, geometry, timing, image quality, and recovery.

Specify the image before selecting the model

The useful requirement is not “good lighting.” It is an image in which the target feature remains distinguishable across expected operating conditions.

Start with the decision. A barcode reader needs clear edges. A cosmetic inspection may need shallow scratches to separate from surface texture. A fill-level check needs a stable boundary through transparent material.

Translate that decision into measurable image properties. Teams can specify minimum contrast, acceptable saturation, blur, shadow coverage, glare, and brightness variation across the field of view.

Camera sensitivity still matters. EMVA 1288 provides standardized methods for measuring and presenting camera and sensor characteristics. Those measurements help teams compare components, but a compliant camera does not guarantee a useful production image.

The optical system includes the target, light source, angle, diffuser, lens, exposure, trigger, enclosure, and surrounding surfaces. The specification must cover that assembled path.

Treat lighting geometry as product configuration

Bright light is not automatically better light.

Basler recommends sufficient illumination without excessive intensity. Too much brightness can destroy detail through saturation, while longer exposure can create motion blur and lower frame rate. Higher gain brightens the image and amplifies noise.

Geometry decides which features become visible. Bright-field illumination can emphasize printed contrast. Dark-field illumination can reveal edges or surface defects. Diffuse light can suppress glare on reflective packaging. Backlighting can produce a clean silhouette.

Capture those choices as versioned configuration:

This is where camera interfaces set the product boundary. Lighting synchronization, trigger latency, cable reach, and control ownership can move with that boundary too.

An undocumented fixture adjusted by hand is not a stable configuration. It is an untracked model input.

Test lighting across the operating envelope

A validation set collected during one shift proves little about the next shift.

Build the test plan around variation the site will actually produce. Include startup, full-speed operation, dirty lenses, aging lights, open doors, reflective lots, darker materials, repositioned parts, and maintenance replacements.

Cognex advises controlling variables such as lighting, working distance, and camera trigger timing when collecting images for deep-learning tools. The same variables must remain represented after launch.

I call the gap between assumed image conditions and monitored image conditions lighting debt. It grows when the model receives more responsibility while illumination remains unmeasured.

Measure performance by condition, not only by an aggregate accuracy score. A product can meet its overall target while failing every glossy package or every image taken near sunrise.

The acceptance report should pair task metrics with image metrics. Record false accepts, false rejects, confidence, contrast, saturation, blur, brightness distribution, and the configuration version used for each run.

Detect image failure before decision failure

The system should know when its input has left the validated envelope.

Add pre-inference checks for clipped highlights, dark frames, blur, obstruction, unexpected color balance, and missing strobes. These checks do not need to diagnose every optical problem. They need to stop bad images from becoming confident product decisions.

Recovery rules should be explicit. The product might retry with a second exposure, switch to another lighting preset, pause the line, route the item for manual review, or notify maintenance.

Keep a small sample of failed and borderline images with their settings. That evidence helps separate model drift from optical drift and gives technicians something better than a red status light.

Lighting hardware also needs maintenance signals. Track strobe counts, controller faults, temperature, intensity changes, cleaning history, and fixture replacement. Revalidate after any change that can alter the image.

Model monitoring begins too late if the photons were already wrong.

A dependable machine vision product defines the image it can judge, proves that image across the operating envelope, and refuses to invent certainty outside it.

Sources

Comments

Comments are powered by GitHub Discussions through Giscus.

← All MitchellCo Notes