Effective augmentation of front opening unified pod filter images

Abstract

The front opening unified pods (FOUPs) are popular carriers in semi-conductor industry, and their automatic optical inspection relies on the sufficiency of images. Filter papers at the bottom of FOUP require frequent inspection, while obtaining their images is time-consuming. This work thus investigates augmenting images of filter papers using conventional interpolation and a recent generative adversarial network with variations, including numbers of epochs, usage of progressive learning, loss functions, and cross frame calibration. Drawbacks of interpolated images are exhibited. Results of sixteen models are quantitatively evaluated using the Fréchet inception distance (FID). The best FID among all models is fid = 0.102, and the distribution of generated images is close to that of real images. A user study is also conducted to investigate the authenticity of generated images. Our experiments reveal that given a limited number of training images, after proper geometric calibration, images synthesized by a generative adversarial network are qualified for data augmentation and substantially benefit industrial image classification tasks.