THE SCRATCH WAS REAL. THE IMAGE MADE IT INVISIBLE. That is a hypothetical scenario, but it describes a genuine machine-vision design risk: the imaging cell must make a defect observable before a classifier can make a defensible decision.

The camera can miss what an operator can see

Imagine a reflective metal component with a shallow scratch. An operator tilts the part and sees the mark immediately. At a fixed camera angle, under fixed illumination, the scratch has almost no contrast. The classifier labels the image PASS. This is an illustrative scenario, not a documented incident or a result obtained from Cognex, Basler, KEYENCE or any other supplier.

The crucial question is not whether the defect exists. It is whether the acquisition system recorded sufficient information to identify it under the conditions in which the decision was made. A model cannot reliably recover physical detail that was not resolved in the image it received.

Lighting is part of the measurement, not a cosmetic accessory

Basler explains how dark-field illumination can reveal surface scratches and edges, while diffuse illumination helps control reflections from shiny or uneven surfaces. Cognex likewise describes the role of lighting geometry, diffuse light and filters in making relevant image features visible. These public explanations support the imaging principle; they do not establish comparative performance for any product.

The right arrangement is application-specific. Material finish, scratch direction, lens, working distance, exposure, motion, field of view and camera mounting all matter. Low-angle illumination that makes one scratch obvious may conceal another feature. Backlighting that is excellent for dimensions may not capture relevant surface texture.

A headline model score cannot replace a production inspection test

An excellent classification score on carefully acquired demonstration images does not prove that the line will expose every specified defect across all acceptable production conditions. A camera upgrade may improve sampling or noise performance, but more pixels cannot compensate for a feature washed out by saturation or inappropriate illumination.

Separate two questions in any acceptance report: could the defect be seen in the raw image, and did the approved inspection decision respond correctly? Record false accepts (defective parts passed) and false rejects (good parts diverted) by defect type, rather than relying only on overall accuracy.

Five acceptance checks worth sharing with the supplier

Use the same representative parts and pre-agreed criteria for each shortlisted integrator. First, document the critical defect taxonomy and the minimum observable feature. Second, capture raw images under normal, low-contrast, reflective and permitted production-speed conditions. Third, retain illumination, lens, exposure, trigger and model/recipe settings. Fourth, measure false accepts, false rejects and inconclusive outcomes separately. Fifth, verify the physical disposition of rejected or uncertain parts and specify what happens when image quality falls outside the approved envelope.

The proposed set is a buyer-side test design, not an industry-standard certification method or an independent supplier ranking. A production-specific sample plan and statistical treatment should be agreed before acceptance.

Make the result usable by quality and maintenance teams

Ask who will maintain lighting geometry and calibration, who may change the inspection recipe, which changes require revalidation and how to restore the last approved configuration. Keep enough traceability to connect the raw acquisition with a part or batch, the decision threshold, the resulting disposition and the software version.

If a demonstration cannot reproduce the evidence behind a PASS, the plant is being asked to trust an outcome rather than verify the inspection process. From image capture to verified quality decisions—that should be the procurement conversation.

Make the next vendor conversation more productive.

Use the free project fit check to generate an initial list of questions for your team.

Open planning tool ↗

Primary sources and scope

The linked primary sources support the stated technical principles or explicitly named published cases. They do not constitute independent product benchmarking. Hypothetical inspection vignette; no tested model, measured defect-escape rate, vendor comparison result or customer incident is claimed.

  1. Basler — Light in Machine Vision
  2. Cognex — The Importance of Lighting
  3. Basler — Lighting in Machine Vision