Date of Award


Degree Type


Degree Name



Electrical and Computer Engineering

First Advisor

Wu, Q. M. Jonathan


3D, Face recognition, Hybrid classifier, Local features, Survey




Face recognition has attracted many researchers’ attention compared to other biometrics due to its non-intrusive and friendly nature. Although several methods for 2D face recognition have been proposed so far, there are still some challenges related to the 2D face including illumination, pose variation, and facial expression. In the last few decades, 3D face research area has become more interesting since shape and geometry information are used to handle challenges from 2D faces. Existing algorithms for face recognition are divided into three different categories: holistic feature-based, local feature-based, and hybrid methods. According to the literature, local features have shown better performance relative to holistic feature-based methods under expression and occlusion challenges. In this dissertation, local feature-based methods for 3D face recognition have been studied and surveyed. In the survey, local methods are classified into three broad categories which consist of keypoint-based, curve-based, and local surface-based methods. Inspired by keypoint-based methods which are effective to handle partial occlusion, structural context descriptor on pyramidal shape maps and texture image has been proposed in a multimodal scheme. Score-level fusion is used to combine keypoints’ matching score in both texture and shape modalities. The survey shows local surface-based methods are efficient to handle facial expression. Accordingly, a local derivative pattern is introduced to extract distinct features from depth map in this work. In addition, the local derivative pattern is applied on surface normals. Most 3D face recognition algorithms are focused to utilize the depth information to detect and extract features. Compared to depth maps, surface normals of each point can determine the facial surface orientation, which provides an efficient facial surface representation to extract distinct features for recognition task. An Extreme Learning Machine (ELM)-based auto-encoder is used to make the feature space more discriminative. Expression and occlusion robust analysis using the information from the normal maps are investigated by dividing the facial region into patches. A novel hybrid classifier is proposed to combine Sparse Representation Classifier (SRC) and ELM classifier in a weighted scheme. The proposed algorithms have been evaluated on four widely used 3D face databases; FRGC, Bosphorus, Bu-3DFE, and 3D-TEC. The experimental results illustrate the effectiveness of the proposed approaches. The main contribution of this work lies in identification and analysis of effective local features and a classification method for improving 3D face recognition performance.

Available for download on Thursday, November 29, 2018