Micron-Level Vision Calibration

I am measuring the shadow of small, irregular plastic parts (1–10 mm) with a machine-vision setup and need to push repeatability below 8 µm. My current pipeline drifts to greater than 10–15 µm on the very same piece, so something in the capture or calibration chain is holding me back.

System snapshot
• Camera: 1280 × 1024 sensor, 4.8 µm pixels
• Optics: zoom lens, field of view 10 × 8 mm
• Method: back-lit silhouette, image processed for length and width

What I need from you
– Review the present imaging and processing flow.
– Build (or refine) a calibration routine that factors pixel size, lens characteristics and stage alignment.
– Tweak image-processing steps—thresholding, edge localisation, sub-pixel fitting, etc.—so the statistical repeatability (σ) stays better than 8 µm across the full FOV.
– Deliver well-commented code or scripts (Python/OpenCV, MATLAB or a tool you suggest) plus a short validation report showing the achieved precision on sample captures.

Acceptance will be based on a test set I supply: the same part measured repeatedly must fall within the 8 µm band, with results and parameters clearly documented.