Diffusion Model-Based Visual Compensation Guidance and Visual Difference Analysis for No-Reference Image Quality...

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Existing No-Reference Image Quality Assessment (NR-IQA) methods struggle with complex distortions and lack interpretable features.
• Leveraging high-level feature information for quality assessment remains a significant challenge in current NR-IQA models.

Purpose of the Study:

• To pioneer the exploration of diffusion models for NR-IQA, addressing limitations of existing methods.
• To develop a novel diffusion model capable of enhancing images with diverse distortions and providing interpretable quality assessment.

Main Methods:

• Designed a novel diffusion model for image enhancement and NR-IQA, utilizing intermediate denoising variables for interpretable guidance.
• Integrated two complementary visual branches to collaboratively leverage high-level visual information for quality evaluation.
• Conducted extensive experiments on seven public NR-IQA datasets to validate the model's performance.

Main Results:

• The diffusion model establishes a clear mapping between image reconstruction and quality scores, effectively guiding the assessment network.
• The proposed model demonstrates superior performance compared to state-of-the-art (SOTA) NR-IQA methods across multiple datasets.
• Achieved enhanced image restoration and more interpretable high-level visual information guidance.

Conclusions:

• Diffusion models offer a promising new direction for NR-IQA, overcoming limitations in handling complex distortions and feature interpretability.
• The developed model provides a robust and effective solution for NR-IQA, outperforming existing SOTA approaches.
• The study highlights the potential of diffusion models in advancing image quality assessment and restoration research.