Project topics and materials
Get Complete Project Material File(s) Now! »
Neural Networks: Key Periods, Models and Algorithms
Despite deep learning has become widely popular only in the recent years, the basic principles and ideas behind ANNs have been developing for more than half a century. Thus, a simplified timelinem with the most significant milestones in the history of ANNs is presented in Figure 2.2.1. In this section, we briefly cover the highlights of the development of ANNs introducing the corresponding central notions, models and algorithms. For a more detailed review of the domain, we invite the reader to refer to the excellent book by Goodfellow et al. [GBC16].
Data Dependent Models
As mentioned in Subsection 2.2.2, the invention of backpropagation resulted in resurgence of the interest in ANNs. In particular, two specific neural models appeared during the 80’s, namely: Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). CNNs are deliberately designed to deal with data which has some spatial topology (i.e. images, videos, sound spectrograms etc.), and are the central instrument for analysis of human visual traits in the present manuscript. Therefore, in this chapter, we devote a separate Section 2.3 for their introduction, while in the rest of this subsection, we focus on RNNs. MLP, which is described in Subsection 2.2.2, is also called feed-forward neural network, because at test stage, the signal passes in a direct manner from the input layer up to the output one. More formally, an oriented connection graph of a feedforward ANN does not contain cycles (cf. Figure 2.2.2). Unlike feedforward ANNs, RNNs contain one or more cycles in their oriented connection graphs. Thus, Figure 2.2.3-(a) illustrates a typical RNN which has the basic structure of an MLP, but each of its hidden neurons is also connected with itself. RNNs are designed to process sequences of data, one element at a time. The easiest way to understand how it works in practice is to refer to Figure 2.2.3-(b) which illustrates unfolding of the RNN from Figure 2.2.3-(a) in time. Thus, an activation x(l+1) j (t +1) of the neuron j of the (l +1)-th layer at time (t +1) depends both on the activations of neurons xl i(t +1) from the previous layer l and on the activation of the same neuron x(l+1) j (t) at the previous time stamp t:
Gender/Age Prediction from Face Images
This section presents a typical pipeline of gender/age prediction systems and makes an overview of
the existing algorithms for addressing the considered problems in Subsections 3.2.1 and 3.2.2, 3.2.3, mrespectively. We end this section by illustrating some potential areas of application for the systems of automatic prediction of gender/age from faces in Subsection 3.2.4.
It is important to highlight that below, we deliberately do not provide the gender recognition and age estimation accuracies of the described approaches from the original articles. The reason is that there is a big discrepancy between the used testing datasets and the various evaluation protocols (i.e. same-dataset, cross-validation, cross-dataset), which makes the resulting scores incomparable between each other. This is especially true for early works on the subjects. Instead, in Tables 5.3.2 and 5.3.3 of Chapter 5, we summarize all scores which were reported on three most popular contemporary gender and age benchmark datasets with well established evaluation protocols.
Standard Pipeline for Gender and Age Prediction from Face Images
Automatic gender recognition and age estimation systems usually follow a typical pipeline which is presented in Figure 3.2.1. It consists of two principle phases, namely: (1) face extraction (or image preprocessing), and (2) face analysis. The objective of the first phase is to extract a face (in a predefined form) from an input image, while the objective of the second phase is to predict gender and age based on the extracted face.
In its turn, face extraction phase is normally composed of three smaller steps. The first one, face detection, is indispensable and is present in all systems of automatic face analysis. Basically, its goal is to detect a face (or faces) in an input image and to output the respective delimiting region (or regions) in the image. Face detection is a classical problem of computer vision with plenty of existing open-source solutions such as [VJ01; Mat+14]. It should be noted that the form of the delimiting face region (square, rectangle, oval etc.) depends on the particular face detector. For example, in this manuscript, we employ a private face detector which is based on [Zha+07] and outputs square-sized face regions.
Table of contents :
Remerciements
Abstract
Résumé
Contents
List of Figures
List of Tables
List of Acronyms
1 General Introduction
1.1 Context and Motivation
1.2 Problems and Objectives
1.3 Contributions
1.4 Organisation of the Manuscript
I State of the Art
2 Deep Learning for Image Analysis and Synthesis
2.1 Introduction
2.2 Neural Networks: Key Periods, Models and Algorithms
2.2.1 Artificial Neuron and Perceptron
2.2.2 MLP and Backpropagation
2.2.3 Data Dependent Models
2.2.4 Deep Learning Revolution
2.3 Convolutional Neural Networks
2.3.1 Typical CNN: Basic Principles and Definitions
2.3.2 Established Training Practices
2.3.3 State-of-the-Art CNN Architectures and CNN Applications
2.4 Deep Generative Models
2.4.1 Overview of Deep Generative Models
2.4.2 Generative Adversarial Networks
2.5 Conclusion
3 Gender/Age Prediction and Editing from Faces
3.1 Introduction
3.2 Gender/Age Prediction from Face Images
3.2.1 Standard Pipeline for Gender and Age Prediction from Face Images
3.2.2 Gender/Age-Aware Feature Extraction
3.2.3 Gender/Age Prediction
3.2.4 Practical Interest
3.3 Gender/Age Synthesis and Editing in Face Images
3.3.1 Face Aging/Rejuvenation
3.3.2 Gender Swapping
3.3.3 Practical Interest
3.4 Conclusion
II Contributions
4 Preliminary Studies
4.1 Introduction
4.2 Study 1: CNN-Learned vs. Hand-Crafted Features
4.2.1 Gender Recognition from Images of Pedestrians
4.2.2 Compared Feature Representations
4.2.3 Experiments
4.2.4 Pedestrian Gender Recognition in Presence of Privacy Protection Filters
4.2.5 Summary of the First Preliminary Study
4.3 Study 2: CNN Architecture for Training from Scratch
4.3.1 Algorithm to Optimize CNN Architecture
4.3.2 Experiments
4.3.3 Summary of the Second Preliminary Study
4.4 Conclusion
5 Gender/Age Prediction from Face Images
5.1 Introduction
5.2 CNN Design and Training Strategy
5.2.1 Previous Studies on Gender and Age Prediction with CNNs
5.2.2 Studied Parameters
5.2.3 Experiments
5.2.4 Summary of the Optimal Design and Training Choices
5.3 Top Performing CNNs for Gender and Age Prediction
5.3.1 Design of the Top Performing CNNs
5.3.2 Benchmark Evaluation
5.3.3 Qualitative Analysis
5.3.4 Top Performing CNNs: Summary
5.4 ChaLearn Competition on Apparent Age Estimation
5.4.1 AAEC Protocol
5.4.2 Proposed Solution
5.4.3 Experiments
5.4.4 AAEC Results
5.5 Conclusion
6 Gender/Age Synthesis and Editing in Face Images
6.1 Introduction
6.2 Face Editing with Conditional Generative Models
6.3 Gender and Age Conditioned Generative Adversarial Network
6.3.1 Design and Training of GA-cGAN
6.3.2 Synthetic Face Manifold
6.3.3 Face Reconstruction via Manifold Projection
6.3.4 Experimental Evaluation of Manifold Projection Approaches
6.3.5 Identity-Preserving Face Reconstruction with GA-cGAN: Summary
6.4 Boosting Cross-Age Face Verification with Age Normalization
6.4.1 Local Manifold Adaptation
6.4.2 Age Normalization
6.4.3 Experiments
6.4.4 GA-cGAN+LMA to Improve Cross-Age Face Verification: Summary
6.5 Conclusion
7 General Conclusion
7.1 Summary of the Contributions
7.2 Limitations and Future Work
7.3 List of Publications
8 Résumé Étendu en Français
8.1 Introduction Générale
8.1.1 Contexte et Motivations
8.1.2 Objectifs
8.1.3 Contributions et Organisation de la Thèse
8.2 Études Préliminaires
8.2.1 Introduction
8.2.2 Étude 1 : Descripteurs Appris par des CNNs vs. Descripteurs Manuellement Conçus
8.2.3 Étude 2 : L’Architecture de CNN pour l’Apprentissage à Partir de Zéro
8.2.4 Conclusion
8.3 Prédiction du Genre et de l’Âge à Partir d’Images de Visages
8.3.1 Introduction
8.3.2 Conception de CNN et Stratégie d’Apprentissage
8.3.3 CNNs les Plus Performants pour la Prédiction du Genre et de l’Âge
8.3.4 La Compétition ChaLearn pour l’Estimation de l’Âge Apparent
8.3.5 Conclusion
8.4 Synthèse et Édition du Genre et de l’Âge dans des Images de Visage
8.4.1 Introduction
8.4.2 GA-cGAN pour la Synthèse et l’Édition du Genre et de l’Âge
8.4.3 Normalisation de l’Âge pour l’Amélioration de la Vérification Faciale
8.4.4 Conclusion
8.5 Conclusion Générale
Bibliography
