Start with the feature you need to see
Write down the defect or dimension that changes the decision. A surface scratch, a missing connector and an unreadable label need different acquisition conditions. Select the field of view and lens around that feature rather than around a convenient camera.
Control variation at acquisition
Treat lighting geometry, part presentation, exposure and the trigger as a single system. Capture samples at the intended motion conditions. Document glare, shadow and position changes that make the feature ambiguous.
Calibration is part of measurement
OpenCV describes intrinsic calibration and lens-distortion correction. These are relevant when mapping image geometry to dimensional results. Record the calibration setup and check stability; the complete measurement method still needs application-specific validation.
Agree on an evidence set
Collect accepted, defective and borderline parts. Keep evaluation samples separate from tuning samples, and include the operational conditions the station will face. Define a hold or manual-review state for images that cannot support a dependable decision.