Abstract

This paper presents a local appearance feature fusion for automatic emotion recognition using Convolutional Neural Network (CNN). The CNN has been known to be a powerful texture feature for facial expression recognition. However, only few approaches utilize the relationship among neighborhood pixels itself. First, CNN is obtained based on two closest vertical and/or horizontal neighborhood pixel relationships. The proposed work is also extended to efficiently handle a large amount of unlabeled data using supervised classification algorithm using modified CNN. At the last stage, ensemble classifiers are trained with a small percentage labeled data and based on the trained model, rest of the unlabeled data is assigned with pseudo-labels.