Convolutional Sparse Autoencoder for Emotion Recognition

Abstract

Emotion recognition is a hot research area in deep learning and computer vision that analyses expressions from both static and dynamic sequences of facial expressions to reveal human emotional states. In recent decades, deep learning approaches have been exhibiting a superior performance on image representation datasets. However, the convolutional neural network (CNN) requires a larger number of labeled datasets for training and accurate classification results. It is always inevitable, whereas unsupervised representation learning models like autoencoder do not require labeled information for training. Meanwhile, it is difficult to infer the feature map when the size of the CNN layer is increased. To address these challenges, this paper introduced a self-supervised deep learning technique called convolutional sparse autoencoder (CSA) which can learn robust features from small data with unlabeled facial expression datasets. Moreover, sparsity is added in the max pooling layer for the feature map which makes the backpropagation optimizer Adam work efficiently for the CSA training; thus, no complicated optimizer is not involved. Finally, the trained convolutional sparse encoder part is combined with the softmax layer for emotion classification. The performance results demonstrate that the proposed approach achieved 98% of accuracy on the CK+ dataset and outperforms various state-of-the-art methods.