A system for automatic face analysis based on statistical shape and texture models [Elektronische Ressource] / Ronald Müller

Sujets

• Informatik

Informations

• Publié par: technische_universitat_munchen
• Publié le: 01 janvier 2008
• Nombre de lectures: 19
• Langue: English
• Poids de l'ouvrage: 8 Mo

Extrait

Lehrstuhl fur Mensch-Maschine-Kommunikation
der Technischen Universit at Munc hen
A System for Automatic Face Analysis
Based on
Statistical Shape and Texture Models
Ronald Muller
Vollst andiger Abdruck der von der Fakult at
fur Elektrotechnik und Informationstechnik
der Technischen Universit at Munc hen
zur Erlangung des akademischen Grades eines
Doktor-Ingenieurs
genehmigten Dissertation
Vorsitzender: Prof. Dr. rer. nat. Bernhard Wolf
Prufer der Dissertation:
1. Prof. Dr.-Ing. habil. Gerhard Rigoll
2. Prof. Dr.-Ing. habil. Alexander W. Koch
Die Dissertation wurde am 28.02.2008 bei der Technischen Universit at Munc hen
eingereicht und durch die Fakult at fur Elektrotechnik und Informationstechnik
am 18.09.2008 angenommen.A System for Automatic Face Analysis
Based on
Statistical Shape and Texture Models
Dissertation
Ronald Muller
Technische Universit at Munc hen
[e-mail]
January 28th 2008Abstract
This dissertation gives an overview and insight in the structure and the scien-
ti c algorithms of a system designed for the automatic analysis of human faces.
Thereby, Face Analysis addresses the goal to extract as much abstract informa-
tion as possible from a face. The applied methods of statistical shape and texture
models base on the idea of Active Appearance Models (AAM). An Appearance
Model for face analysis describes the variations in shape and texture of human
faces derived from a careful selection of photographs showing di erent persons
with di erent facial expressions and head poses in various lighting conditions de-
pending on the speci c focus of the analysis. During the analysis of a human face
within a video or a picture, the Appearance Model is used to re-synthesize this
face as optimal as possible. Apart from an introduction to AAMs with a uni ed
mathematical notation, this document describes the various optimizations and
modi cations on several steps of the basic algorithm.
While the recognition and interpretation of faces is comparatively lightweight
for the human visual cortex, this task requires computer vision approaches of high-
est computational complexity. Thus, this thesis not only ghts the challenge of
most accurate face analysis, but also the di culties of building up an integrated,
fully automatic software system which provides a high computational e ciency
plus techniques for the extensive exploitation of modern standard hardware.
The evaluations compare the di erent developed algorithms with respect to
the quality of the re-synthesized face, computational complexity, and pattern
recognition tasks, such as the determination of e.g. the gender, age, head pose,
and facial expression of a person.
Acknowledgments Apart from the o cial bodies of the Technische Univer-
sit at Munc hen, especially Professor Dr.-Ing. habil. Gerhard Rigoll, my special
thanks go to all the students who conducted their Master, Diploma, and Bachelor
Theses, their Interdisciplinary Projects, and Seminar Presentations with me. I
very much enjoyed the collaboration with them and I am happy having built a
fruitful network of young students and professionals. This work would not have
been such successful without the indescribable diligence, e orts, and skills of Ralf
Nikolaus and Michael Geisinger. They deserve the greatest thank and respect for
their contribution. Eventually, I thank Karin Hammerschmid in devotion for her
considerateness, relief, and footing.
Trademarks Trademarks appear throughout this document without any trade-
mark symbol; they are the property of their respective trademark owner. There is
no intention of infringement; the usage is to the bene t of the trademark owner.Contents
Abstract i
Contents iii
1 Introduction 1
1.1 The Challenge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The Background . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 FEASy { a FacE Analysis System . . . . . . . . . . . . . . . . . 3
1.4 The Thesis in a Nutshell . . . . . . . . . . . . . . . . . . . . . . . 5
2 A Multi-Threading Framework for Signal Processing Systems 7
2.1 Conditions for the development of Pro Systems . . . 8
2.2 Other works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Requirements of a Software Framework for High-Performance Sig-
nal Processing Systems . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Concepts of MMER Lab . . . . . . . . . . . . . . . . . . . . . . . 11
2.4.1 Software Architecture . . . . . . . . . . . . . . . . . . . . . 14
2.4.2 Design Decisions . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Application Examples and Evaluation . . . . . . . . . . . . . . . . 15
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 Object Localization with AdaBoost Variants on Haar- and
Gabor-Wavelet Features 19
3.1 Haar-like and Gabor-Wavelet features . . . . . . . . . . . . . . . . 20
3.1.1 Haar-like features . . . . . . . . . . . . . . . . . . . . . . . 20
3.1.2 Gabor-Wavelets . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Feature Selection and Classi cation with AdaBoost . . . . . . . . 24
3.2.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2.2 The Standard AdaBoost Algorithm . . . . . . . . . . . . . 25
3.2.3 Gentle AdaBoost . . . . . . . . . . . . . . . . . . . . . . . 26
3.2.4 Weak classi ers . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2.5 Cascaded AdaBoost Classi cation . . . . . . . . . . . . . . 29
3.3 Evaluation of Localization Performance . . . . . . . . . . . . . . . 30
3.3.1 Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.2 Head and Eye Localization Results . . . . . . . . . . . . . 31iv CONTENTS
3.3.3 Feature selection . . . . . . . . . . . . . . . . . . . . . . . 32
3.3.4 Localization Performance . . . . . . . . . . . . . . . . . . . 33
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4 The Theory of Active Appearance Models 37
4.1 Preparation of Training Data . . . . . . . . . . . . . . . . . . . . 38
4.1.1 Alignment and Normalization of Landmarks . . . . . . . . 38
4.1.2 Warping . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.1.3 Normalization of Textures . . . . . . . . . . . . . . . . . . 41
4.2 Generation of an Appearance Model . . . . . . . . . . . . . . . . 43
4.2.1 Shape Model . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2.2 Texture Model . . . . . . . . . . . . . . . . . . . . . . . . 44
4.2.3 Combined Model . . . . . . . . . . . . . . . . . . . . . . . 44
4.3 Coe cient Optimization . . . . . . . . . . . . . . . . . . . . . . . 45
4.3.1 Objective Function . . . . . . . . . . . . . . . . . . . . . . 46
4.3.2 O ine Prediction . . . . . . . . . . . . . . . . . . . . . . . 48
4.3.3 Numerical Estimation of the Jacobian Matrix . . . . . . . 49
4.3.4 Iterative Optimization . . . . . . . . . . . . . . . . . . . . 51
5 Derivatives and Advancements of Active Appearance Models 53
5.1 A Survey on Active Appearance Models and Variants . . . . . . . 53
5.2 Appearance Models based on NMF . . . . . . . . . . . . . . . . . 55
5.2.1 Data Modeling with Non-Negative Matrix Factorization . . 56
5.2.2 Generation of Appearance Models with NMF . . . . . . . 68
5.2.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.3 Online Optimization of AAM Coe cients . . . . . . . . . . . . . . 72
5.3.1 Gradient Descent . . . . . . . . . . . . . . . . . . . . . . . 73
5.3.2 Grid Sampling . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.3.3 Nelder-Mead or Simplex Optimization . . . . . . . . . . . 76
5.3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.4 GPU-Accelerated Active Appearance Models . . . . . . . . . . . . 79
5.4.1 Warping . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.4.2 Coe cient Optimization . . . . . . . . . . . . . . . . . . . 86
5.4.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.5 Evaluation Measures for the Quality of AAM Re-synthesis . . . . 89
5.5.1 Dataset Annotation . . . . . . . . . . . . . . . . . . . . . . 89
5.5.2 Quality Measures . . . . . . . . . . . . . . . . . . . . . . . 90
5.5.3 Evaluation of Quality Measures . . . . . . . . . . . . . . . 93
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6 Application of Active Appearance Models to Face Analysis 95
6.1 Classi cation Based on Results of the AAM Optimization . . . . . 96
6.1.1 Classi cation based on class speci c AAMs . . . . . . . . . 96
6.1.2 Statistical classi cation based on AAM coe cients . . . . 97
6.1.3 Support Vector Machines . . . . . . . . . . . . . . . . . . . 98CONTENTS v
6.1.4 N-fold Cross-Validation . . . . . . . . . . . . . . . . . . . . 99
6.2 Image Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.2.1 The AR Database . . . . . . . . . . . . . . . . . . . . . . . 100
6.2.2 The NIFace1 Database . . . . . . . . . . . . . . . . . . . . 101
6.2.3 The FG-NET Aging Database . . . . . . . . . . . . . . . . 101
6.2.4 The MMI Face Database . . . . . . . . . . . . . . . . . . . 102
6.3 Gender Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.3.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.3.2 Results . . . . . . . . . . . . . . .