Combining Local and Global Measures for DIBR-Synthesized Image Quality Evaluation

Area of Science:

• Computer Vision
• Image Processing
• Multimedia Systems

Background:

• Depth-Image-Based-Rendering (DIBR) is crucial for 3D video applications like 3D TV and Free Viewpoint Video (FVV).
• DIBR-synthesized images often contain distortions impacting the viewer experience.
• Existing quality evaluation methods (Full-Reference and No-Reference) struggle with DIBR-specific artifacts.

Purpose of the Study:

• To develop a novel No-Reference (NR) quality evaluation method for DIBR-synthesized images.
• To address limitations of existing methods in assessing DIBR-related distortions.
• To enable the design of perceptually friendly FVV systems.

Main Methods:

• Proposed an NR method focusing on geometric distortions (disocclusion, stretching) and sharpness.
• Geometric distortions were assessed via local similarity and region comparison.
• Sharpness was measured using scale-invariant properties by comparing images with downsampled versions.

Main Results:

• The proposed method effectively evaluates perceptual quality by combining distortion and sharpness scores.
• Experimental results demonstrate superiority over prevailing Full-Reference (FR) and No-Reference (NR) metrics.
• Outperformed competing methods in effectiveness (except APT) and significantly in implementation time.

Conclusions:

• The developed NR method provides a robust and efficient solution for DIBR image quality assessment.
• It offers a practical tool for enhancing visual quality in 3D video applications.
• The method's speed and accuracy make it suitable for real-time FVV systems.