Data augmentation by using gans and image transformation in facial emotion classification

Abstract

We propose different approach combinations including GANs and image transformation in facial emotion classification especially for FER-2013 dataset. Our goal is not to reach the highest accuracy for this dataset, but to compare and verify the data augmentation effects for several main GAN methods such as Vanilla GAN, CGAN, DCGAN and BiGAN. We used these GAN models to generate fake images and put them to our train dataset. And the data together was through image transformation to train in the CNN model that we designed to perform facial emotion classification. Base on the result when we used some GANs (BiGAN/DCGAN) and image transformation, the accuracy of the model indeed has been improved. The best combination data augmentation approaches in this paper are using image transformation and DCGAN.