Data Encoding Visualization Based Cognitive Emotion Recognition with AC-GAN Applied for Denoising

Abstract

Emotion is a subjective, conscious experience when people facing internal or external stimuli. This paper addresses the problem that affective computing is difficult to be put into real-world practical fields intuitively, such as emotion disease diagnosis and so on, due to the non-intuitive data representation. In view of the fact that people's ability to understand two-dimensional images is much higher than that of one-dimensional data, we use Markov Transition Fields to visualize time series signals. MTF images represent the first order Markov transition probability along one dimension and temporal dependency along the other. Besides, with the limitation of experimental equipment and individual differences among volunteers, noise is inevitable. We apply AC-GAN to remove noisy pixels within high dimension and acquire high resolution images before making classification. Then we use Tiled Convolutional Neural Networks on 2 real world datasets to learn high-level features from MTF images. The classification results of our approach are competitive with the state-of-the-art approach. This method makes the visualization based emotion recognition become possible, which is beneficial for the application of cognitive robots or in the medical fields, such as depression and other psychological problems diagnosis, and can help doctors and patients understand the condition more intuitively.