Towards Realistic Facial Expression Recognition

Abstract

Automatic facial expression recognition has attracted significant attention over the past decades. Although substantial progress has been achieved for certain scenarios (such as frontal faces in strictly controlled laboratory settings), accurate recognition of facial expression in realistic environments remains unsolved for the most part. The main objective of this thesis is to investigate facial expression recognition in unconstrained environments. As one major problem faced by the literature is the lack of realistic training and testing data, this thesis presents a web search based framework to collect realistic facial expression dataset from the Web. By adopting an active learning based method to remove noisy images from text based image search results, the proposed approach minimizes the human efforts during the dataset construction and maximizes the scalability for future research. Various novel facial expression features are then proposed to address the challenges imposed by the newly collected dataset. Finally, a spectral embedding based feature fusion framework is presented to combine the proposed facial expression features to form a more descriptive representation. This thesis also systematically investigates how the number of frames of a facial expression sequence can affect the performance of facial expression recognition algorithms, since facial expression sequences may be captured under different frame rates in realistic scenarios. A facial expression keyframe selection method is proposed based on keypoint based frame representation. Comprehensive experiments have been performed to demonstrate the effectiveness of the presented methods.