No-Reference Quality Assessment for View Synthesis Using DoG-based Edge Statistics and Texture Naturalness.

Area of Science:

• Computer Vision
• Image Processing
• Human Visual System Modeling

Background:

• View synthesis for free-viewpoint video relies on texture and depth images.
• Distortions in synthesized views arise from both image acquisition/processing and rendering (DIBR).
• Existing quality metrics fail to assess distortions from the entire view synthesis pipeline.

Purpose of the Study:

• To develop a no-reference quality index for synthesized views that addresses distortions from all stages.
• To propose a metric that accurately evaluates the overall quality of the view synthesis process.

Main Methods:

• A novel no-reference quality index, Synthesized view quality index using DoG-based Edge statistics and Texture naturalness (SET), is presented.
• The metric utilizes Difference-of-Gaussian (DoG) representation to capture edge and texture characteristics, mimicking human visual perception.
• Multi-scale DoG images are computed, and orientation-selective statistics and texture naturalness features are extracted.
• A random forest regression model is employed to learn the quality model from these features.

Main Results:

• The proposed SET metric effectively captures edge degradation and texture unnaturalness common to both synthesis stages.
• Experimental results validate SET's superior performance over state-of-the-art metrics on established view synthesis databases.
• SET demonstrates robustness in evaluating distortions across the entire view synthesis pipeline.

Conclusions:

• The SET metric provides a more accurate and comprehensive evaluation of synthesized view quality compared to existing methods.
• This approach enhances the assessment of free-viewpoint video quality by considering all distortion sources.
• SET offers a valuable tool for researchers and developers in the field of view synthesis.