Emotion recognition using machine learning techniques for robots

Abstract

Emotion Recognition is one of the classification tasks in the computer vision, carrying interactive communication between human and machines. This project aims to set up an emotion recognition system in a household robot. The recognition system is realized by balancing the factors in terms of hardware constraint and recognition accuracy. More specifically, the household robot is supposed to conduct a few tasks but within a limited 2GB memory space, therefore, the software must be designed memory-compactly. As for the real-time test using a webcam, the model first tries to capture faces in a video frame by using the HOG feature face detector, then it applies several preprocessing techniques such as Gaussian blurring, adaptive histogram equalization and mean- subtraction to the face and then sends it to the pre-trained smaller AlexNet model for the recognition task. The test accuracy of the model reaches 0.71 and the highest recognition rate reaches 0.90 for a happy face on FER-2013 test set.