Defect detection in interferometers through supervised regression-based machine learning

Context

Manufacturer of microscopes and optical measuring instruments, precision manufacturing. Defects in interferometer images are to be detected and quantified automatically — repeatably, across shifts and inspectors.

Problem

Visual inspection was inconsistent and quality costs were rising. A sufficiently large, cleanly annotated dataset of real defects for classic training did not exist.

What we built

A machine-learning solution for detecting and quantifying defects in images from specialised microscopic devices, built for repeatable, consistent calculations: supervised regression, trained on a synthetically generated dataset of ideal images with artificially added defects.

Key engineering decision

Supervised regression with automatically generated training data. Instead of merely classifying defects, the model estimates defect areas and their size. The training dataset was generated automatically — ideal images combined with synthetically added defects. That made training independent of the availability of real defect images and provided exactly known ground truth for every training sample.

Result

• 98.5 % confidence score
• Defective areas are identified reliably, their sizes calculated precisely
• Repeatable results — independent of shift and inspector