Robustness comparison between the capsule network and the convolutional network for facial expression recognition

Abstract

As an important part of human-computer interactions, facial expression recognition has become a popular research topic in computer vision, pattern recognition, artificial intelligence and other fields. With the development of deep learning and convolutional neural networks, research on facial expression recognition has also made considerable progress. Because facial expressions vary in real environments, such as rotation, shifting, brightness changes, partial occlusion and noise with different intensities, research on the robustness of facial expression recognition is very important. A capsule network consists of capsules, which are groups of neurons, and these capsules can learn posture information through the dynamic routing mechanism. The length of a capsule represents the existence probability, and each neuron in a capsule represents posture information (e.g., position, size, orientation or a combination of these properties). Therefore, in this study, the robustness of the emerging capsule network (CapsNet) is comprehensively compares with that of the traditional convolutional neural network (CNN) and fully convolutional network (FCN) in facial expression recognition tasks. The simulation results based on the Cohn-Kanade (CK+) databases show that the capsule network is more robust than the other networks. Therefore, the capsule network has significant advantages over the other networks in facial expression recognition task in complex real-world environments.