Exploring generative techniques for defect image synthesis

Details

Supervisors

Lee, John Aldo

Degree label

Master [120] : ingénieur civil en science des données, à finalité spécialisée

Abstract

In recent years, there has been a significant increase in the use of machine learning applications to address a wide range of problems. However, training these models requires substantial amounts of data. While much of this data is relatively easy to gather, certain types, such as defect images, remain challenging to collect. This is particularly true in production lines, where conforming objects vastly outnumber defective ones. This document explores techniques for generating defect images by leveraging the similarities between conforming and defective objects to artificially create defect data for machine learning. To assess the usefulness of this generated data, classification neural networks will be trained using the augmented data. This will help evaluate how effectively the generated data improves the accuracy of the classifier.