Dynamic Facial Feature Learning by Deep Evolutionary Neural Networks

Abstract

Facial Action Coding System is a comprehensive and anatomical system which could encode various facial movements by the combination of basic AUs (Action Units), and makes the emotion categories much wider. Recently, deep learning has been shown its superiority on recognition tasks. Despite the powerful feature learning ability of deep learning, there are still several problems remained. Firstly, a large amount of training data is needed to fully extract features and avoid overfitting. Secondly, the parameters optimization of deep neural network is complex, and the direct guidance of the results is insufficient. In this paper, a spatiotemporal self-learning method is designed by evolutional deep neural network model, and spatial augmentation is utilized to deal with the two problems facing in practical application. The proposed method is conducted on AUs analysis task which is important for emotion identification. The 3D convolutional neural network which could learn dynamic facial features from AUs image sequences is optimized automatically for the topology and hyper-parameters by evolutional scheme. Extensive experiments demonstrated the effectiveness of EVONET (Deep Evolutionary Neural Networks) on the facial databases over alternative methods, including 3DCNNs (3D Convolutional Neural Networks), and several convolutional neural network based models.