Master of Engineering by Research
University of Wollongong. School of Electrical, Computer and Telecommunications Engineering
Li, Peiyao, Adaptive feature extraction and selection for robust facial expression recognition, Master of Engineering by Research thesis, University of Wollongong. School of Electrical, Computer and Telecommunications Engineering, University of Wollongong, 2010. https://ro.uow.edu.au/theses/3268
Facial expression is one of the means humans convey their emotional states. Accurate recognition of facial expressions should therefore lead to advances in human computer communication and applications in robotics and mimetic games. This project investigates a novel approach to recognise facial expressions from static images. We propose a method for face alignment to address the localisation errorin existing face detection methods, through eye detection and face verification.In the proposed facial expression recognition approach, fixed and adaptive 2-D filters are combined in a hierarchical structure. The fixed filters are used to extract primitive features such as edges and directions, whereas the adaptive filters are trained to extract more complex and subtle facial features for classification. Both types of filters are non-linear and they are based on the biological mechanism of shunting inhibition. Linear classifier and Support Vector Machines are both used as classifiers in the facial expression recognition system.
To improve the system performance, two feature selection algorithms are proposed to select salient features for classification. One is based on symmetric Kullback-Leibler divergence, the other is based on mutual information. The proposed approaches are evaluated on the JAFFE database, which has seven types of facial expressions: anger, disgust, fear, happiness, neutral, sadness and surprise. The proposed face detection and alignment method can correctly align and crop all the images from the database. The system with linear classifiers achieves a facial expression classification rate of 95.9%, while system based on SVMs has a higher recognition rate of 96.7%. Using an efficient feature selection method, the system can achieve the best performance with a smaller feature subset. This facial expression recognition system compares favourably with several existing techniques tested on the same database.