Hyperparameter optimization for convolutional neural network by opposite-based particle swarm optimization and an empirical study of photomask defect classification

Abstract

Automated optical inspection (AOI) technology has been widely used for defect detection in various industries including wafer maps and TFT-LCD. Although the implementation of AOI significantly reduce the efforts of manual visual tests, most of the automated were leaving out defect classification for manual inspection. Previous studies indicate that various technologies including statistical methods or machine learning techniques can be used to classify the defect images. Among them, convolutional neural networks (CNN) bring out attention to classification problems. Indeed, determining the suitable choices of the critical hyperparameters is crucial for classification performances. However, most related studies determine the hyperparameters by testing the performance among a few combinations. Despite some previous studies integrating metaheuristics with CNN to improve the selection of hyperparameters, only a few studies focus on AOI defect classification by using CNN with hyperparameter optimization. This study proposed an opposite-based particle swarm optimization with CNN (OPSO-CNN) to deal with the AOI defect classification of semiconductor photomask while considering the hyperparameter optimization of the proposed CNN model, in which opposite-based initialization mechanism was used to generate the initial particles to increase the diversity of the searching space. An empirical study was conducted to evaluate the performances of the proposed OPSO-CNN by the benchmarks. The results have shown the practical viability of the proposed model.