Automatically Design Lightweight Neural Architectures for Facial Expression Recognition

Abstract

Facial expression recognition (FER) is a popular direction researched in the field of human-computer interaction. Recently, most of the work in the direction of FER are with the help of convolutional neutral networks (CNNs). However, most of the CNNs used for FER are designed by humans, and the design process is time-consuming and highly relies on the domain expertise. To address this problem, some methods are proposed based on neural architecture search (NAS), which can automatically design neural architectures. Nevertheless, those methods mainly focus on the accuracy of the recognition, but the model size of the designed architecture is often large, which limits the deployment of the architecture on devices with limited computing resources, such as mobile devices. In this paper, a novel approach named AutoFER-L is proposed for automatically designing lightweight CNNs for FER. Specifically, the accuracy of recognition and the model size are both considered in the objective functions, thus the resulting architectures can be both accurate and lightweight. We conduct experiments on CK+ and FER2013, which are popular benchmark datasets for FER. The experimental results show that the CNN architectures designed by the proposed method are more accurate and lighter than the handcrafted models and the models derived by standard NAS.