Зено Бассель Синтез изображений лиц на основе генеративных методов машинного обучения с применением к распознаванию лиц

Краткое описание

Зено Бассель Синтез изображений лиц на основе генеративных методов машинного обучения с применением к распознаванию лиц
ОГЛАВЛЕНИЕ ДИССЕРТАЦИИ
кандидат наук Зено Бассель
Реферат

Synopsis

Introduction

Chapter 1 Analytical overview of methods for face image synthesis, setting goals and objectives of the study

1.1 Historical overview for face image synthesis

1.1.1 Statistical models and subspace representation

1.1.2 Geometry modeling

1.1.3 Basic augmentation methods

1.1.4 Learning methods

1. 2 Theorical background

1.2.1 Discriminative and generative models

1.2.2 Deep neural networks

1.3 Performance metrics

1.3.1 Face recognition performance metrics

1.3.2 Image quality assessment metrics

1.4 State of the art

1.4.1 Explicit 2D/3D modeling for face image synthesis

1.4.2 End to end methods for face image synthesis using generative adversarial network

1.4.2.1 GANs for image-to-image translation

1.4.2.2 Cross-domain image-to-image translation

1.4.2.3 Multi-domain image-to-image translation

1.4.2.4 Face image frontalization using generative adversarial network

1.4.2.5 Face image rotation using generative adversarial network

1.4.2.6 Face image animating using generative adversarial network

1.4.2.7 face image inpainting using generative adversarial network

1.5 Summary

Chapter 2 Development of methods and algorithms for face image generation with the given attributes and preserving the identity

2.1 Comparative analysis between Cross-Domain and Multi-Domain image-to-image translation

2.1.1 Motivation

2.1.2 The proposed evaluation metric to compare image-to-image translation models

2.1.3 The proposed loss functions for improving preservation the identity

2.1.4 Experiments

2.2 Developing generative method for learning identity and pose disentanglement (IP-GAN)

2.2.1 Motivation

2.2.2 IP-GAN network architecture

2.2.3 The learning algorithm of the disentangled identity and pose representation

2.3 Developing multi-functional generative method for controlled synthesis of face image (CtrlFaceNet)

2.3.1 Motivation

2.3.2 CtrlFaceNet network architecture

2.3.3 The learning algorithm

2.4 Visual data augmentation methodology

2.5 Summary

Chapter 3 Software implementation of the proposed methods and experimental results

3.1 Implementation of the developed method for learning identity and pose disentanglement (IP-GAN)

3.1.1 Experiments and discussion

3.1.1.1 Dataset

3.1.1.2 Implementation details

3.1.1.3 Features visualization

3.1.1.4 Face pose transformation

3.1.1.5 Random face generation

3.1.1.6 Identity similarities

3.1.1.7 Face image quality quantitative results

3.1.1.8 Face verification

3.1.1.9 Face identification task

3.1.1.10 Performance evaluation on large-scale dataset of unconstrained face still images

3.2 Implementation of the multi-functional developed method for controlled synthesis of face image (CtrlFaceNet)

3.2.1 Experiments and discussion

3.2.1.1 Dataset

3.2.1.2 Implementation Details

3.2.1.3 Controlled face image generation with a driving face image

3.2.1.4 Comparison results of controlled face image generation

3.2.1.5 Face geometry morphing

3.2.1.6 Face image inpainting

3.2.1.7 Face verification task using feature extractor trained on augmented dataset

3.2.1.8 Face identification task

3.3 Software implementation of facial images synthesis based on generative machine learning methods

3.3.1 Software development and programming tools

3.3.2 Implementation of algorithms for generating face images

3.4 Summary

Conclusion

List of abbreviations and conventions

References

List of Own Publications

List of Figures

List of Tables

Appendix A. Copies of Author's Publications