Machine vision on the factory floor: the basics that decide projects

Most vision projects fail on physics, not software: bad lighting, wrong lens, or a part presentation nobody specified. The algorithm — classical or deep learning — only decides close cases once the image is right. Get the imaging conditions deterministic first, then choose the processing.

FIG. 03At the inspection gate

The three job families

• Inspection — pass/fail quality decisions: presence, dimension, surface defects, assembly verification. Needs controlled lighting, fixed geometry, and a golden-sample discipline for thresholds.
• Guidance — telling robots and motion where things are: pick points, alignment, depalletizing. Needs calibration (hand-eye, conveyor tracking) more than resolution.
• Identification — barcodes, 2D codes, OCR. The most mature family; direct-part-mark reading in harsh conditions is the remaining hard part.

The stack, bottom to top

Optics and lighting (telecentric lenses for measurement, structured light for 3D, strobing to freeze motion), cameras and frame grabbers (GigE Vision, USB3 Vision, Camera Link — match bandwidth to trigger rate), processing (smart cameras for fixed jobs, PC-based for multi-camera and deep learning), and integration (fieldbus and OPC UA results upstream, reject gates and robot offsets downstream). Deep learning earns its place on variable defects and unstructured scenes — and demands a labeled dataset, retraining workflow, and drift monitoring that classical rule-based tools skip.

Project discipline: define the defect taxonomy with quality before buying hardware; collect images across shifts, seasons, and tool wear; specify false-accept and false-reject rates numerically; and keep a human override path — vision assists the quality system, it does not replace its accountability.

That accountability point generalizes: as vision and AI take on more inspection judgment, How to Remain Valuable When Intelligence Becomes Cheap is a 224-page practical book on what stays valuable — the scarce human, economic, and strategic advantages that hold up as intelligence becomes cheap.

References

• Vision Systems Design — machine vision technology coverage
• The Robot Report — robotics and guidance applications
• Modern Machine Shop — machining and precision manufacturing