Expression recognition based on multiple feature fusion-based convolutional neural network

Abstract

The process of traditional convolutional neural network to deal with human expression features is often one-way, and it cannot make full use of the features of each layer in the convolutional network to retain valuable facial expression features. This paper proposes an improved convolutional neural network structure with a multi-feature fusion module to optimize facial expression recognition. Within this structure, three convolutional layers are divided into one area module. For each area module, the input of each convolutional layer is composed of the outputs of the previous two convolutional layers. At the same time, a dense connection structure is added between two different area modules to realize feature reuse, thereby relying on feature diversity to improve the accuracy of facial expression recognition. Our experimental results on the CK+ and FER2013 databases show that compared with existing convolutional neural networks, the method gives higher accuracy and lower structure complexity.