DeepDC: Deep Distance Correlation as a Perceptual Image Quality Evaluator.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep neural networks excel at image quality assessment but require pixel alignment.
• Existing methods struggle with natural images and texture similarity due to alignment constraints.

Purpose of the Study:

• To develop a novel full-reference image quality assessment (FR-IQA) model that overcomes spatial alignment limitations.
• To leverage deep statistical similarity for robust image quality evaluation.

Main Methods:

• Utilized pre-trained deep features without fine-tuning or spatial co-location.
• Employed distance correlation, a statistical measure, to quantify feature similarity.
• Derived a closed-form solution for distance correlation using deep double-centered distance matrices.

Main Results:

• The proposed FR-IQA model demonstrates superior performance and robustness across various benchmarks.
• Achieved state-of-the-art results in optimizing texture synthesis and neural style transfer.
• Validated through extensive experimental evaluations.

Conclusions:

• The novel FR-IQA model effectively assesses image quality without pixel-level alignment.
• Deep statistical similarity offers a powerful alternative for FR-IQA and related tasks.
• The method shows significant promise for practical image analysis and synthesis applications.