P‐120: The Solution for Imbalanced Image Data of Mass Production

Abstract

AMOLED is the core technology in display industry which is made through the glass substrate process. As the AMOLED's resolution becomes higher, the glass becomes thinner, and the inspection system for securing the glass process reliability becomes an issue in the manufacturing process. Currently, the inspection is conducted manually on random samples, thus it is impossible to inspect all the defects. For full inspection, therefore, there is a need for a high‐performance AI model which has been trained by a balanced training set. However, imbalanced data has been raised as a big issue in the mass production field. In this study, the solution for the imbalanced image data training model has been developed to automatically classify and determine the crack defects, thereby enhancing the detection consistency of the defects. This paper applies redesigned DeepSMOTE [1] based on deep learning for an inspection system Glass‐Edge‐Crack‐Detection (GECD) process. This enhanced DeepSMOTE is capable of generating high‐quality, artificial images that can enhance minority classes and balance the training set. The effect of this technology is to solve the imbalanced data and improve the performance of the GECD based on deep learning. Through this paper, the imbalanced image data is solved and the foundation of the solution for imbalanced image data technology applicable to mass production is established. In the test, finally we got the results that ‘zero leakage' (0/411) and ‘NG recall 100%' performance of the model. We propose the enhanced DeepSMOTE in the OLED mass production