Classification of Emotions Using a 2-Channel Convolution Neural Network

Abstract

This paper focuses on facial emotion recognition with the help of a 2-channel convolution neural network. The first channel in the CNN model contains a standard network and the other layer is made up of local features of the images. Information from both channels is connected to a fully connected layer which is used for classification using a log loss function. The designed scheme is tested on FER2013 dataset. The results obtained from the designed system are compared with standard CNN and SVM using accuracy, precision and Fl score. The results reveal that the 2-channel CNN is far better than the other models in terms of recognizing emotions correctly. It also revealed that providing local feature descriptors of images using CNN helps in improving classification by a significant margin in low-resolution images.