Image Quality Assessment: Exploring Joint Degradation Effect of Deep Network Features via Kernel Representation...

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep network-based full-reference image quality assessment (FR-IQA) models typically analyze features pairwise.
• Existing methods overlook correlations among features from the same source, limiting comprehensive quality assessment.

Purpose of the Study:

• To propose a novel dual-branch framework for FR-IQA that captures joint degradation effects among deep network features.
• To enhance perceptual awareness in image quality assessment by integrating feature self-similarity analysis with pairwise comparisons.
• To develop a training-free method for deriving image quality scores that can guide perceptual image enhancement.

Main Methods:

• A dual-branch framework incorporating Kernel Representation Similarity Analysis (KRSA) and pairwise comparisons using Mean Absolute Error (MAE).
• KRSA branch analyzes feature self-similarity matrices, while the second branch performs standard pairwise feature comparisons.
• A training-free logarithmic summation of both branches generates the final image quality score.

Main Results:

• The proposed method demonstrates enhanced perceptual awareness by considering joint degradation effects.
• Experiments across 10 datasets validate the method's efficacy in diverse IQA scenarios.
• The approach successfully identifies and differentiates appealing content deformations.

Conclusions:

• Integrating KRSA with pairwise comparisons improves the perceptual accuracy of FR-IQA models.
• The dual-branch framework is adaptable to various network architectures, offering flexibility.
• The developed method provides a valuable tool for guiding perceptual image enhancement and understanding image quality.