Robust facial feature localization in frontal views

Abstract

In this thesis, we focus on the reliable detection of facial fiducial points in frontal face images, such as eye, eyebrow and mouth corners. The proposed algorithm aims to improve automatic landmarking performance in challenging realistic face scenarios subject to high-valence facial expressions and occlusions. In order to extract facial fiducial points, we explore the potential of several feature modalities, namely, Gabor Wavelet Transform (GWT), Independent Component Analysis (ICA), Non-negative Matrix Factorization (NMF) and Discrete Cosine Transform (DCT), both singly and jointly. The multitude of landmark candidates is associated via fusion techniques. We show that the selection of the highest scoring face patch as the corresponding landmark is not always the best, but that there is considerable room for improvement with the cooperation among several high scoring candidates and also using a graph-based post-processing method. The developed methods are tested on Bosphorous, JAFFE (Japanese Females Facial Expression Database) and BioID face image databases. We also present comparative results with "Elastic Bunch Graph Matching" algorithm. The performance of each method and each conducted experiment is discussed separately.