Impact of Data Dimensionality Reduction on Neural Based Classification: Application to Industrial Defects

摘要

A major step for high-quality optical surfaces faults diagnosis concerns scratches and digs defects characterisation. This challenging operation is very important since it is directly linked with the produced optical component's quality. To complete optical devices diagnosis, a classification phase is mandatory since a number of correctable defects are usually present beside the potential "abiding" ones. Unfortunately relevant data extracted from raw image during defects detection phase are high dimensional. This can have harmful effect on behaviors of artificial neural networks which are suitable to perform such a challenging classification. Reducing data dimension to a smaller value can however decrease problems related to high dimensionality. In this paper we compare different techniques which permit dimensionality reduction and evaluate their possible impact on classification tasks performances.