Which output is needed?
Classification assigns an image label, detection locates objects with boxes, and segmentation defines pixel-level regions. Checking whether a box is present and measuring a scratch are different tasks. First specify the output an operator needs and how that output will be used.
Planning image quality
Stable mounting, a suitable viewing angle and controlled lighting are sensible starting points. When glare, shadows, vibration or motion blur appear, review image acquisition before trying to compensate in software. Select samples from expected product and environmental variations as well as ideal conditions.
Decisions and uncertainty
A visual result is an input to a business process. Define a path for uncertain samples: human review, another capture or a safe wait. Acceptance testing should record false alarms and missed cases separately on representative samples. This evaluates value against operational needs rather than an impressive demonstration.