Handling class imbalance and multiple inspection objectives in design of industrial inspection system

Abstract

A large number of real-world machine learning and data mining problems have several crucial issues that need to be solved, such as class imbalance and multiple objectives; these issues cannot be easily overcome with many learning methods. This study proposes an efficient approach to address the issues and applies the approach in the design of a machine vision system for a real-world problem: inline inspection of surface defects on glass substrates of thin-film transistor liquid crystal displays (TFT-LCDs). The three major steps in the machine vision system are: (1) selective extraction of defect features from images using 2-dimemnsional wavelet decomposition, (2) training cost-sensitive classifiers to handle class imbalance, and (3) use of ensemble techniques to achieve multiple manufacturing objectives. When applied to an industrial case study, the achieved performance shows that using the proposed approach in defect inspection of TFT-LCD glass substrates is a viable alternative to manual inspection.