Analyzing the Influence of Different Activation Functions Based on Deep Learning Model for Facial Expression Recognition

Abstract

Facial expressions are an important channel for people to communicate their emotions. To more accurately recognize people's facial expressions, researchers are constantly exploring the possibilities of convolutional neural networks. For convolutional neural network models, many factors can have a significant impact on the performance, including the structure and parameters. In this paper, it analyze the impact of different activation functions on the deep learning model of facial expression recognition with the FER-2013 dataset, compare the advantages and disadvantages between traditional and new activation functions, and finally build a deep learning model of facial expression recognition with better performance. In addition to the baseline CNN model, the paper also analyzes the performance of famous deep learning models such as ResNet, VGG and Inception, from which the best-performing baseline CNN model is selected to explore the impact of different activation functions. The results show that the GELU activation function-based facial expression deep learning model has the best performance and the highest recognition accuracy among the activation functions ReLU, L-ReLU/P-ReLU, Swish, etc. Compared with the deep learning model with the traditional ReLU activation function, the facial expression deep learning model based on GELU activation function constructed in this paper approximately improves the accuracy by 1%.