Be a hairstylist - editing hair for face images using generative adversarial network
Hair editing has long been a complicated problem in computer graphics. This is due to it being challenging to simulate the complex nature of every single strand of hair. Even though there are many conventional 3D hair modelling methods that can create a realistic visual appearance, they are often af...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Final Year Project |
| Language: | English |
| Published: |
Nanyang Technological University
2023
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/166146 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | Hair editing has long been a complicated problem in computer graphics. This is due to it being challenging to simulate the complex nature of every single strand of hair. Even though there are many conventional 3D hair modelling methods that can create a realistic visual appearance, they are often affected by a major limitation which is consuming many computing resources due to the involvement of 3D matters. As such, this can slow down the process considerably. Recently, there has been a surge of research studies and interest in Generative Adversarial Networks (GANs) and GANs can be a potential solution to this issue. Researchers have been studying ways to further enhance the GANs’ performance and applicability. One such application is image-to-image translation which transforms an image to another domain, whose result contains the style in the new domain while retaining the content of the original domain. The 3D hair modelling methods can be replaced by using image-to-image translation and this eliminates the need to model for each hair strand. Models are instead, needed to be trained to represent the mapping between two domains. As such, the use of image-to-image translation with GANs is a feasible solution for solving hair editing challenges in computer graphics. In this project, to address the problem of hair editing in computer graphics, we will be using GAN-based image-to-image translation method along with an array of other image segmentation and image synthesis techniques such as face parsing and Image Synthesis with Semantic Region-Adaptive Normalization (SEAN) to produce a high-fidelity facial image with the edited hairstyle. |
|---|