Universal blind image quality assessment metrics via natural scene statistics and multiple kernel learning

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Universal blind image quality assessment (IQA) is crucial for image processing when ground truth is unavailable or distortion types are unknown.
• Current state-of-the-art blind IQA algorithms rely on natural scene statistics (NSS) but have limitations in accurately capturing non-Gaussianity (NG) and local dependency (LD), and do not utilize exponential decay characteristics (EDC).

Purpose of the Study:

• To develop advanced universal blind IQA metrics by integrating NG, LD, and EDC, addressing limitations of existing NSS-based methods.
• To incorporate multiple kernel learning (MKL) to handle the heterogeneous properties of different image features.

Main Methods:

• Proposed two new universal blind IQA models: NSS global scheme and NSS two-step scheme.
• Exploited NG using wavelet coefficients' marginal distribution and measured LD using mutual information.
• Incorporated EDC features directly and employed MKL for feature similarity measurement with diverse kernels.

Main Results:

• The proposed metrics exhibit remarkable consistency with human perception across various distortions.
• Both developed models significantly outperform existing representative universal blind IQA algorithms.
• The metrics also surpass standard full-reference quality indexes in performance.

Conclusions:

• The novel universal blind IQA metrics effectively utilize NG, LD, and EDC, offering improved performance over existing methods.
• The integration of MKL enhances the ability to measure feature similarity, contributing to better quality prediction.
• These metrics provide a robust solution for blind image quality assessment in practical image processing applications.