Title Page
Abstract
Contents
Chapter 1. Introduction 10
Chapter 2. Related Works 15
2.1. Single Image Deblurring 15
2.1.1. General Deblurring 15
2.1.2. Face Deblurring 16
2.2. Generative Adversarial Networks 17
2.2.1. Conditional GANs 18
2.2.2. U-Net based GANs 18
Chapter 3. Proposed Method 20
3.1. Semantic-Aware Pixel-Wise Projection Discriminator 21
3.2. Discriminator training 23
3.3. Generator training 25
Chapter 4. Experiments and Results 28
4.1. Experimental Details 28
4.1.1. Datasets 28
4.1.2. Implementation Details 29
4.1.3. Evaluation Metrics 29
4.2. Comparisons on MSPL dataset 30
4.3. Comparisons on real blurred images 34
4.4. Execution time and Face Verification 35
4.5. Ablation Study 36
Chapter 5. Conclusions 39
Reference 40
Table 4.1. Quantitative Comparisons on MSPL testset. 30
Table 4.2. Comparison with recent SFID methods for average run time, model parameters and verification accuracy. 35
Table 4.3. Effectiveness of different components of SAPPGAN on the MSPL-Center testset. 36
Figure 1.1. Comparison of discriminator architectures. 11
Figure 3.1. Overall architecture of the proposed face deblurring framework. 21
Figure 4.1. Qualitative comparison on MSPL-Center testset. 31
Figure 4.2. Qualitative comparison on MSPL-Random testset. 32
Figure 4.3. Qualitative comparison on Real-Blur test set. 34