Artificial Intelligence based Facial Emotion Recognition with Deep Neural GAN Augmentation

Facial recognition system in computer vision is widely used in human-computer interface and clinical diagnostics. Nonetheless, achieving high accuracy in this task still poses a challenge due to issues of inadequate amount and imbalance of data to train the models as well as issues of generalization and robustness. The findings of this research propose the use of Generative Adversarial Network (GAN) data augmentation for face emotion classification specifically for this approach. The proposed strategy is based on the generation of artificial ‘emotional’ images on the face using Deep Neural Self Attention GANs (DNSAGANs). This way, to capture all the variations of facial expressions the GAN takes a large number of images and thus makes the augmented data richer and more realistic. This larger data set is then used to train a deep learning architecture that performs the face emotion classification thereby enabling the network to learn enhanced discriminative features. To collect face photos the FER2013 dataset is used. The training process, entrained the generator and discriminator to minimize the mean squared error while the discriminator was simultaneously trained to robustly distinguish generated images from real ones. Altogether, the proposed Adam optimizer was more accurate than Adagrad and Adadelta for all epochs. The percentage of accuracy is about $88.64 \%$ for 100 epochs using the proposed optimizer techniques.