Facial Expression Recognition Based on An Improved Data Augmentation Method and A Multichannel Convolutional Neural Network Model

Abstract

With the development of artificial intelligence and human-computer interaction, the recognition and analysis of facial expressions have recently received increasing attention. Considering the excellence of the convolutional neural network (CNN) model in the field of image recognition, this paper applies the CNN model to recognize facial expressions. The deep neural network is inseparable from large-scale data. In the real world, it is expensive to obtain large scale data sets with labels. A reliable method of expanding the number and diversity of the samples must be found. The generative adversarial networks (GAN) model can generate more images and solve the imbalance of data volume. However, the experiments indicated that the image sharpness generated by the traditional GAN model was insufficient. The model was challenging to train stably and prone to over-fitting or under-fitting. Given these situations, this paper optimized the structure of the GAN model by combining the advantages of the DCGAN and StarGAN models, and improved the training strategy. Moreover, a new multichannel CNN model is proposed and applied in this paper to obtain more comprehensive facial expression features. The experimental results reveal that the proposed method significantly improves the performance of the evaluated baseline facial expression recognition methods.