Developing a pixel difference prediction model for face aging

Face-Aging, a branch of image generation research, aims to age or rejuvenate facial images to a specified target age while preserving individual identity traits. This technology is significant in various sectors, including public safety, entertainment, facial recognition, and skin analysis. Currentl...

Full description

Saved in:
Bibliographic Details
Main Author: Liang, Xuheng
Other Authors: Alex Chichung Kot
Format: Thesis-Master by Coursework
Language:English
Published: Nanyang Technological University 2024
Subjects:
Online Access:https://hdl.handle.net/10356/174254
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Face-Aging, a branch of image generation research, aims to age or rejuvenate facial images to a specified target age while preserving individual identity traits. This technology is significant in various sectors, including public safety, entertainment, facial recognition, and skin analysis. Currently, the domain is dominated by GAN-based Face-Aging techniques. However, these methods show limitations, especially in maintaining consistent facial expressions and in the overall quality of the generated images. This dissertation focuses on addressing identity and expression inconsistency problems in Face-Aging. By utilizing real face data and an open-source method, a Longitudinal Aging Dataset was synthesized, redefining Face-Aging as an image translation task. A novel Face-Aging Pixel Difference prediction model based on U-Net was developed, guided by the FRAN framework principles. In order to improve the consistency, we incorporate Haar wavelet skip connections and Age Layer Positional Encoding. The model demonstrated controlled aging effects, improved image quality, and preserved identity and expression consistency on the validation set. Nonetheless, challenges persist in real-face data test scenarios. This study’s key contributions are the development of an enhanced Face-Aging approach and a comprehensive assessment of existing technologies, setting a foundation for future advancements in model optimization and application expansion.