Blind Image Quality Assessment Based on High Order Statistics Aggregation.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Current blind image quality assessment (BIQA) methods rely on handcrafted features unsuitable for text/graphics or learning-based features requiring extensive training.
• Existing methods face limitations in efficiency and applicability due to large codebooks or complex updating procedures.

Purpose of the Study:

• To propose a novel, general-purpose BIQA method utilizing high order statistics aggregation (HOSA).
• To develop an efficient and accurate BIQA approach requiring only a small codebook.

Main Methods:

• High Order Statistics Aggregation (HOSA): Extracts local image features, constructs a small codebook (100 codewords) using K-means clustering, and aggregates high-order statistics (mean, variance, skewness).
• Feature Extraction: Utilizes local normalized image patches on a regular grid.
• Quality Prediction: Employs Support Vector Regression (SVR) to map aggregated features to subjective opinion scores.

Main Results:

• The proposed HOSA method demonstrates highly competitive performance against state-of-the-art BIQA methods.
• Evaluated on ten diverse image databases with simulated and realistic distortions, HOSA shows robust quality assessment capabilities.

Conclusions:

• HOSA offers an effective and efficient general-purpose blind image quality assessment solution.
• The method overcomes limitations of existing BIQA approaches by requiring only a small codebook and avoiding complex training procedures.