Use this URL to cite or link to this record in EThOS:
Title: Linear subspace methods in face recognition
Author: Nguyen, Hieu
ISNI:       0000 0004 2713 5386
Awarding Body: University of Nottingham
Current Institution: University of Nottingham
Date of Award: 2011
Availability of Full Text:
Access from EThOS:
Access from Institution:
Despite over 30 years of research, face recognition is still one of the most difficult problems in the field of Computer Vision. The challenge comes from many factors affecting the performance of a face recognition system: noisy input, training data collection, speed-accuracy trade-off, variations in expression, illumination, pose, or ageing. Although relatively successful attempts have been made for special cases, such as frontal faces, no satisfactory methods exist that work under completely unconstrained conditions. This thesis proposes solutions to three important problems: lack of training data, speed-accuracy requirement, and unconstrained environments. The problem of lacking training data has been solved in the worst case: single sample per person. Whitened Principal Component Analysis is proposed as a simple but effective solution. Whitened PCA performs consistently well on multiple face datasets. Speed-accuracy trade-off problem is the second focus of this thesis. Two solutions are proposed to tackle this problem. The first solution is a new feature extraction method called Compact Binary Patterns which is about three times faster than Local Binary Patterns. The second solution is a multi-patch classifier which performs much better than a single classifier without compromising speed. Two metric learning methods are introduced to solve the problem of unconstrained face recognition. The first method called Indirect Neighourhood Component Analysis combines the best ideas from Neighourhood Component Analysis and One-shot learning. The second method, Cosine Similarity Metric Learning, uses Cosine Similarity instead of the more popular Euclidean distance to form the objective function in the learning process. This Cosine Similarity Metric Learning method produces the best result in the literature on the state-of-the-art face dataset: the Labelled Faces in the Wild dataset. Finally, a full face verification system based on our real experience taking part in ICPR 2010 Face Verification contest is described. Many practical points are discussed.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available
Keywords: QA 75 Electronic computers. Computer science ; TA Engineering (General). Civil engineering (General)