Low Complexity Deep Learning for Mobile Face Expression Recognition

Abstract

The problem of Face Expression Recognition (FER) remains a challenging one due to variations in illumination and pose as well as partial occlusion of the face. Deep neural networks have been increasingly applied to this problem and have achieved excellent recognition results, especially on challenging datasets such as FER2013. However, the trend has been towards more complex networks to increase performance. In this paper, we develop a low complexity model, and we experiment with a variety of parameters to determine the performance of these models on the FER2013 dataset relative to the complexity of the models. We show that we are able to obtain an accuracy of 70.86% on the test FER images which approximately matches the winning entry to the FER2013 competition but our model is 5 times smaller in size. We show that we are able to reduce the model size 5 times more, resulting in a model with fewer than 500,000 parameters, and still maintain an excellent accuracy of 68.43% which would make this model ideal for resource constrained environments.