Lightweight Deep Learning Model for Facial Expression Recognition

Facial expression recognition (FER) is one of the fundamental cornerstones for many applications such as driver fatigue monitoring, social robotics, and medical treatment. It is of great challenge to recognize facial expression with high accuracy because the settle features are not easily captured. Deep learning expects to play an essential role in tackling this challenge and improving facial expression recognition accuracy. Although existing complicated deep learning models can achieve high accuracy, the computational cost is too high for resource constrained devices such as internet of thing devices. In this work, we propose a lightweight deep learning model based on MobileNetV2 and Inception, to reduce computational cost while maintaining relatively high accuracy. Specifically, the proposed model uses an Inception convolutional neural network (CNN) to extract features from inputs, and the backbone bottlenecks to compress the model and learn function efficiently; and the CNN block to expand the learned features in bottleneck layers and feed the fully-connected layer for classification. The proposed model is highly efficient, and with small size, that can be deployed on devices equipped with limited computational resources. We conduct experiments, and the results demonstrate that the proposed lightweight deep learning model can achieve relatively high accuracy for expression recognition and reduce the computational cost significantly.