| Author: | Liu, Qiyun |
| Title: | Deep learning-based face image inpainting using Res-WGAN |
| Degree: | M.Sc. |
| Year: | 2023 |
| Department: | Department of Electrical and Electronic Engineering |
| Pages: | 74 pages : color illustrations |
| Language: | English |
| Abstract: | Face inpainting is the process of restoring a defective face image by means of a computer. The face image is restored according to certain rules with the help of the field pixel information of the defective field to achieve a restoration effect that the human observer cannot recognize whether it is a defective image or not. Traditional image inpainting includes old image restoration, obscuration removal, video information recovery, etc. However, images restored with traditional techniques often suffer from blurring or artifacts. When dealing with large areas of defective images, traditional techniques are very ineffective. When it comes to face images which are rich in semantic information, neither traditional image restoration methods nor the use of information diffusion principles to predict defective areas are sufficient for restoration. This paper proposes a face restoration algorithm based on Res-WGAN, which is innovative in that it can obtain background information from the deep network to complete face restoration in an unsupervised manner. This dissertation improves on the common problems of traditional restoration algorithms with the following main work: (a) A study of the theoretical model of generative adversarial networks and its training methods. The network is a classic work of generative models, but also has many shortcomings such as training instability, convergence difficulties, gradient spread, etc. Later, scholars proposed improved models such as WGAN and DCGAN to be able to use GAN for image processing and to solve the above-mentioned problems that tend to arise. (b) Introducing a residual network with a spectral normalization algorithm based on WGAN to make the network satisfy Lipschitz continuity without changing the parameter matrix structure and enhance the training capability of the network. (c) In this paper, a loss function consisting of a trinity of contextual loss, perceptual loss, and Wasserstein loss is defined to train the network. We first select 50,000 randomly selected images from CelebA as input to the network, then input the images to be repaired into the network to generate fake images, and select the best fake images based on the innovative loss function, which are then appended to the defective images to achieve the purpose of face image inpainting. We finally verified the restoration effect by objective metrics PSNR and SSIM, which improved by xx and could obtain clearer visual restoration effect compared with using traditional GAN and DCGAN. |
| Rights: | All rights reserved |
| Access: | restricted access |
Files in This Item:
| File | Description | Size | Format | |
|---|---|---|---|---|
| 8272.pdf | For All Users (off-campus access for PolyU Staff & Students only) | 2.42 MB | Adobe PDF | View/Open |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13869