Development of perceptual image quality database and assessment for AI generated facial images
In recent years, advancement in generative Text-to-image (T2I) models have made great strides, contributing to the emergence of AI-Generated Facial Images (AGFIs), with diverse applications in various fields including entertainment, education and research. However, different AGFIs possess large qual...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Final Year Project |
| Language: | English |
| Published: |
Nanyang Technological University
2024
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/175321 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | In recent years, advancement in generative Text-to-image (T2I) models have made great strides, contributing to the emergence of AI-Generated Facial Images (AGFIs), with diverse applications in various fields including entertainment, education and research. However, different AGFIs possess large quality variance, demanding the need for image quality benchmark models that are consistent with human subjective ratings. Therefore, this project examines various popular T2I AI-Generated Image (AGI) models and leveraged different prompts to generate AGFIs. Through rigorous subjective assessments, a new human perception-based quality database for T2I AGFIs, titled AGFI-500, is established. Subsequently, a benchmark experiment is conducted on this database. Using different popular Image Quality Assessment (IQA) metrics from existing literature to evaluate the performance of current IQA methodology in the domain of scoring perceptual quality of AGFIs. Lastly, a novel solution for No-Reference Image Quality Assessment (NR-IQA) termed Generated Facial NR-IQA (GFNR-IQA), is proposed as an alternative to existing techniques for assessing individual AGFIs and generating perceived quality score. |
|---|