A New Approach for Automatic Face Emotion Recognition and Classification Based on Deep Networks

Recognition of human emotions has been a challenging topic in field of human-computer interaction. In order that there is more natural interaction between human and computer, the computer must be able to recognize, distinguish and respond to human emotions. Also Automated Face Expression Recognition (FER) is still continuing to be a challenging and concerning problem in Computer Vision. In spite of all the efforts being made in the evolution of various methods for FER, the present methods lack popularity when it comes to unseen images or pictures captured in wild settings. This paper tends to design an artificially intelligent system capable of emotion recognition through facial expressions of unknown people. The network in this paper consists of three convolutional layers each followed by max pooling and ReLU. The network is trained on FER2013 dataset and tested on RaFD dataset thus giving a wide range of training images to the network, so that it can overcome the basic problem of recognition of unknown faces. The pertinence of the final model is depicted in a live video application that can instantaneously return users emotions based on their facial posture. The accuracy obtained by this method was 68%, which is better than the previous state-of-the-arts methods. The results provide an important insight on the significance of using different datasets for training and validation.