Machine Vision Project Planning for Facilities and Production

What machine vision is good at

• Checking presence or absence of components, labels and caps.
• Measuring dimensions, gaps and positions.
• Detecting surface defects such as scratches, dents and contamination.
• Reading printed text, barcodes and 2D codes such as lot numbers and expiry dates.
• Verifying assembly steps and orientation.
• Counting items on conveyors or in trays.

Define the problem precisely

• A defect catalog with images of each defect type and borderline examples.
• Acceptance criteria agreed with quality: what size or severity makes a part a reject?
• Line speed, parts per minute and how parts are presented to the camera.
• Product variation: colors, finishes, sizes and packaging changes.
• The current inspection method and how many defects escape it.
• The relative cost of rejecting good parts versus shipping bad ones.

Test feasibility first

Capture images of good and defective parts with candidate cameras, lenses and lighting. A simple test: can a person clearly see the defect in the captured image? If not, change the optics or lighting before thinking about software. Check that the smallest defect covers enough pixels to be detected reliably, and that exposure time is short enough to avoid motion blur at line speed, using strobed lighting if needed.

Hardware decisions

Rule-based vision or deep learning

Data collection and labeling

• Collect images across shifts, batches, suppliers and seasons.
• Label with quality experts using written definitions, and check that labelers agree.
• Expect few defect examples; plan anomaly detection or targeted collection of defects.
• Keep a held-out test set that is never used for training.
• Version datasets and models together.

Integration with the line

• Signals to PLCs for reject mechanisms, with confirmation that rejected parts actually left the line.
• Results and images logged to quality or manufacturing execution systems.
• An operator screen that shows results and the reason for each reject.
• Operational technology network segmentation for cameras and vision computers.
• A latency budget that fits the time between image capture and the reject point.

Validate before switching over

• Test with reference sets of known good and known defective parts.
• Measure false rejects and escaped defects against the acceptance criteria.
• Run the same parts through several times to check repeatability.
• Run the system in shadow mode beside the current inspection before relying on it.

Running it in production

• Monitor reject rates and image quality, since lights age and lenses get dirty.
• Schedule cleaning and calibration checks.
• Retrain models through change control, with the held-out test repeated each time.
• Keep spare cameras, lenses and lights, and train operators and maintenance staff.
• If cameras capture people, for example in safety monitoring, handle the images as personal information.