No-reference high-dynamic-range image quality assessment based on tensor decomposition and manifold learning.

Area of Science:

• Computer Vision
• Image Processing
• Perceptual Computing

Background:

• Full-reference image quality assessment (IQA) methods have limited practical applications.
• High-dynamic-range (HDR) image quality assessment remains a challenge, especially without reference images.

Purpose of the Study:

• To propose a novel no-reference IQA method for HDR images.
• To achieve objective quality prediction scores that closely match human visual perception.

Main Methods:

• Tensor decomposition to extract three feature maps (color, structure) from HDR images.
• Manifold learning on the primary feature map to capture inherent geometric structures.
• Multi-scale feature extraction (manifold structure and contrast features).
• Support vector regression for aggregating features and predicting quality scores.

Main Results:

• The proposed no-reference method outperforms existing full-reference and no-reference IQA methods.
• Demonstrated superiority over the full-reference HDR-VDP-2.2 method on the Nantes database.
• Achieved higher consistency with human visual perception compared to other methods.

Conclusions:

• The developed no-reference IQA method offers a practical solution for assessing HDR image quality.
• The approach effectively leverages feature extraction and machine learning for accurate quality prediction.
• The method shows significant promise for real-world applications requiring objective HDR image evaluation.