A conveyor carries ring-shaped parts under an overhead camera and ring light. A deterministic synthetic image is thresholded into foreground pixels, and the measured area and center-hole brightness determine the accept or reject decision. You compare physical defects with lighting-induced errors.
• An indexed inspection conveyor, camera, lens and focus barrel, ring illumination, image and binary-mask monitors, and a reject diverter with collection bins. • Controls for dark-object threshold (40-180 gray level), illumination (0.4-1.4 relative), defocus (0-3 pixels), area tolerance (5-40 %) and part population (good, missing sector, blocked hole, or all of one type). • Readouts of measured foreground pixels, area deviation, center-hole mean brightness, accepted and rejected parts, and decisions disagreeing with the true condition. • Experiments: correct lighting and threshold, underexposed good parts, and a loose area tolerance.
The nominal image is an annulus of radius 10-20 pixels centered at (31.5, 31.5). A pixel is foreground if its gray level is below the threshold and area is the foreground count. A part is accepted if |area − nominal area|/nominal area ≤ tolerance and the center-hole mean exceeds the threshold. Good is an intact annulus, a notch is a missing 45° sector and a plug fills the center, with one decision per 2 s.
This is a deterministic synthetic image-processing demonstration, not a trained AI model or real camera calibration. It has no lens distortion, motion blur, noise, connected-component filtering or metrology scale correction, and the decision uses only area and hole brightness, so it has known blind spots.
No. Lighting and threshold settings can cause false rejection; underexposed good parts can fall below the threshold and be rejected.
No. It applies explicit threshold, area and hole-brightness rules to actual pixels.
With a loose area tolerance a missing-sector part can pass, since the decision uses only area and hole brightness.
From the simulated part generator, which is what lets the lab count decisions that disagree with the true condition.