Vision Inspection Accuracy: The Four Factors That Decide Detection Performance

Vision Inspection Accuracy: What Really Determines Whether a Sorting Machine Catches Your Defects

Two sorting machines can carry identical camera counts and price tags yet perform completely differently on the same parts. Detection capability is a system property — optics, lighting, handling and software together. Understanding the four factors below will make you a much sharper buyer of vision inspection equipment.

1. Optical resolution vs real precision

Resolution math is simple: field of view divided by sensor pixels gives µm per pixel, and reliable measurement needs 3-5 pixels across the smallest feature. But real precision also depends on telecentric optics (no perspective error as parts shift), calibration discipline, and sub-pixel algorithms. Ask suppliers for a Gage R&R style repeatability study on your critical dimension — not just the sensor datasheet.

2. Lighting is half the machine

Every defect type has a lighting geometry that reveals it: backlight for silhouettes and dimensions, coaxial light for flat surface scratches, low-angle ring light for dents and burrs, dome light for shiny curved surfaces. A machine that catches your cracks uses the right light, not just more cameras. This is why sample testing beats specification comparison — the supplier must actually solve your defect list.

3. Part presentation and handling

The best camera cannot inspect what it cannot see consistently. Glass dial machines give stable, repeatable presentation for small parts; vibratory feeding must orient parts without jamming or damaging them; rotation stations expose cylindrical surfaces. Feeding reliability determines real-world throughput far more than camera frame rate — watch the machine run an hour with your parts before you believe any parts-per-minute number.

4. Software: rules, tolerances and the AI question

Classic rule-based vision measures dimensions and thresholds deterministically — auditable and stable, ideal for dimensional judgment. Deep-learning classification helps on cosmetic defects with natural variation (stains, texture anomalies) where rules over-reject. Modern systems, including Unitecho platforms, combine both: rules for geometry, trained models for appearance — with every reject image archived so quality engineers can audit each decision.

5. Proving performance before purchase

The only meaningful acceptance test: seed a known defect set into good parts, run it blind, and count escapes and false rejects. Repeat at full speed. Insist this test is written into the contract with agreed limits. Reputable machine builders welcome it — it is how we prove the machine and how you defend the investment internally.

6. Working with Unitecho

Send sample parts and your defect list; we run feasibility imaging, build the inspection recipe, and demonstrate escape/false-reject performance on video before shipment. Machines ship with recipe management, defect image archiving and SPC export, and our engineers support remote recipe tuning after installation.