基于GAN和CNN模型的人脸画像合成方法

发布时间：2021-05-20 15:52

　　异质图像合成是指对各种不同来源人脸图像之间进行转化与合成,比如不同光照变化下相机拍摄的人脸照片、画家手绘的素描画像、软件合成的人脸照片以及红外成像设备采集到的红外图像。近年来,应用在数字娱乐领域的异质图像合成以及应用在执法领域中的素描画像的合成与识别受到了极大的关注。人脸画像合成主要是指通过输入的照片生成相应的素描画像,主要通过一些合成方法对画像-照片之间的复杂映射关系进行建模,并利用所学习到的映射关系来合成输入照片所对应的素描画像。在理想情况下,合成的画像或照片图像应该保留更多的外观纹理,并且越逼真越好。这样,它才具有良好的视觉感知质量和较高的识别精度。鉴于此,本文基于深度学习方法开展关于人脸画像合成和识别的相关研究。首先,本文对一些典型的人脸画像合成方法进行了全面的回顾和比较。目前尚未见关于人脸画像合成方法的实验对比和分析的研究。鉴于合成过程与训练模型相关联,现有方法可以分为两大基本类型:数据驱动的方法和模型驱动的方法。根据人脸画像合成过程,数据驱动方法,又称为基于样本的方法,一般包含四个步骤:图像块表示、近邻选择、权重计算以及图像块拼接。而模型驱动方法则直接学习人脸照片和画像之间... 

【文章来源】：西安电子科技大学陕西省 211工程院校 教育部直属院校

【文章页数】：114 页

【学位级别】：博士

【文章目录】：
ABSTRACT
摘要
Nomenclature
ACRONYMS
Chapter 1 Introduction
    1.1 Heterogeneous Image Synthesis
        1.1.1 Digital Entertainment
        1.1.2 Law Enforcement
        1.1.3 Examples of FSS Application
    1.2 History and State-of-the-art Methods
        1.2.1 Conventional State-of-the-art Methods
        1.2.2 Deep Learning-Based Methods
    1.3 Structure of the Thesis
        1.3.1 Main Contents of This Thesis
        1.3.2 Organization of the Thesis
Chapter 2 Comparative Study on Some Typical FSS Methods
    2.1 Introduction
    2.2 Data-Driven Methods
        2.2.1 Subspace Learning-Based Methods
        2.2.2 Sparse Representation-Based Methods
        2.2.3 Probabilistic Graphical Model-Based Methods
    2.3 Model-Driven Methods
        2.3.1 Linear Model-Based Methods
        2.3.2 Nonlinear Model-Based Methods
    2.4 Experiments and Analysis
    2.5 Data Preparation
        2.5.1 Database Specification
        2.5.2 Representative Methods Settings
    2.6 Data Analysis
        2.6.1 Face Sketch Synthesis
        2.6.2 Image Quality Assessment
        2.6.3 Face Sketch Recognition
    2.7 Chapter Summary
Chapter 3 A Novel FSS Approach with GAN and CNN Models
    3.1 Introduction
    3.2 Related Work
        3.2.1 Previous FSS Methods
        3.2.2 Deep Learning-Based FSS Methods
    3.3 Proposed Method
        3.3.1 Fully Functional Framework
        3.3.2 Coarse Estimation Through GAN Model
        3.3.3 Fine Estimation Through CNN Model
        3.3.4 Algorithm For Proposed Method
    3.4 Chapter Summary
Chapter 4 Performance Evaluation of FSS Methods
    4.1 Introduction
    4.2 General Image Quality Assessment (IQA)
        4.2.1 Subjective Evaluation
        4.2.2 Objective Evaluation
        4.2.3 Structural Similarity Index Metric (SSIM)
    4.3 Weighted SSIM For Synthesized Sketch Quality Assessment
        4.3.1 Gaussian Approach
        4.3.2 Human Specified Technique
    4.4 Sketch Recognition For Performance Evaluation of FSS
        4.4.1 Principle Component Analysis (PCA)
        4.4.2 Eigenface Technique
        4.4.3 Linear Discriminant Analysis (LDA)
        4.4.4 Null space Linear Discriminant Analysis (NLDA)
    4.5 Chapter Summary
Chapter 5 Experimental Results and Analysis
    5.1 Introduction
    5.2 Parameter Specification
        5.2.1 For GAN Model
        5.2.2 For CNN Model
    5.3 Face Sketch Synthesis
    5.4 Camera Captured Face Photos Database
    5.5 Objective Image Quality Assessment
    5.6 Face Recognition
    5.7 Chapter Summary
Chapter 6 Concluding Remarks and Future Work
    6.1 Concluding Remarks
    6.2 Future Work
References
Acknowledgements
Author Profile



本文编号：3198027