No-Reference Image Quality Assessment Leveraging GenAI Images.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning methods show promise for image quality assessment but struggle with limited real-world data and poor generalization.
• Existing challenges in image quality assessment stem from the scarcity of annotated, real-world training datasets.

Purpose of the Study:

• To propose a novel no-reference image quality assessment (NR-IQA) method that leverages generative AI (GenAI) images.
• To address the limitations of data scarcity and poor generalization in current NR-IQA models.

Main Methods:

• Utilized GenAI images as reference images, employing a cold diffusion model to generate distorted images across four distortion types.
• Labeled distorted images using a full-reference model to construct a large-scale pre-training dataset for NR-IQA model development.
• Integrated a Multi-scale Cross Attention Block (MCAB) and a Scale Simple Attention Module (SSAM) to enhance feature representation by extracting multi-scale information.

Main Results:

• The proposed method demonstrated state-of-the-art (SOTA) performance across eight public image quality assessment databases.
• The developed large-scale pre-training dataset significantly improved the building of NR-IQA models.
• Enhanced feature representation through MCAB and SSAM modules proved effective in predicting image quality.

Conclusions:

• The proposed GenAI-based NR-IQA method effectively overcomes data limitations and achieves SOTA performance.
• The novel dataset generation and feature extraction techniques offer a promising direction for future NR-IQA research.
• The study facilitates the development of more robust and generalizable image quality assessment models.