Quality Prediction on Deep Generative Images

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks, particularly generative adversarial networks (GANs), excel at realistic image generation.
• Assessing the perceptual quality of generated images is crucial for monitoring and controlling image compression processes.
• Current image quality assessment algorithms struggle with GAN-generated content, particularly in textured areas and at high compression levels.

Purpose of the Study:

• To develop a novel image quality assessment (IQA) model specifically designed for generative images produced by GANs.
• To address the limitations of existing IQA algorithms in evaluating GAN-generated content.

Main Methods:

• A multi-stage parallel boosting system was employed for the new GAN IQA model.
• The model utilizes structural similarity features and measurements of statistical similarity.
• A subjective GAN image quality database was created, including distorted GAN images and human quality ratings.

Main Results:

• The proposed GAN IQA model demonstrated superior prediction performance on generative image datasets.
• The model also showed strong performance on traditional image quality datasets.
• Experimental results validate the effectiveness of the naturalness-based approach.

Conclusions:

• The developed naturalness-based IQA model provides accurate quality predictions for GAN-generated images.
• This advancement is significant for the reliable assessment and control of generative image compression.
• The new model offers a robust solution for evaluating image quality in the context of generative AI.