Machine Vision Defect Inspection
Deep-learning based visual inspection prototype that flags surface defects on production parts.
Overview
A camera-based inspection prototype that classifies parts as pass/fail from images and exposes results to the line controller.
Problem
Visual inspection by operators was slow, subjective and inconsistent across shifts.
Solution
Trained a convolutional classifier on labelled defect images and wrapped it in a Python service that returns a verdict per part, with a confidence threshold routed to a reject actuator.
Engineering process
Image acquisition and lighting setup, dataset labelling, augmentation, model training and evaluation, then integration testing against the controller.
Results
Consistent classification on the validation set and a clear confidence threshold for borderline parts requiring manual review.