Model-Based Referenceless Quality Metric of 3D Synthesized Images Using Local Image Description

Area of Science:

• Computer Vision
• Image Processing
• Virtual Reality

Background:

• Emerging 3D technologies like virtual reality (VR) and augmented reality (AR) present new challenges for video quality assessment.
• Free Viewpoint Video (FVV) is a key next-generation technology, but its synthesis using Depth Image-Based Rendering (DIBR) in blind environments necessitates reliable quality evaluation.
• Existing metrics fail to accurately reflect human judgment due to geometric distortions inherent in DIBR.

Purpose of the Study:

• To propose a novel, reliable, real-time, referenceless quality metric for DIBR-synthesized images.
• To address the limitations of current assessment metrics in capturing geometric distortions.
• To enhance the accuracy of blind quality evaluation for FVV.

Main Methods:

• Utilizing autoregression (AR)-based local image description for DIBR-synthesized images.
• Calculating reconstructed error between DIBR images and AR-predicted images to capture geometric distortions.
• Incorporating visual saliency to refine the proposed blind quality metric.

Main Results:

• The proposed AR-based method accurately captures geometric distortions in DIBR images.
• The metric, modified by visual saliency, shows improved performance.
• Experimental results demonstrate the superiority of this no-reference method over existing full-, reduced-, and no-reference models.

Conclusions:

• The novel referenceless quality metric effectively evaluates DIBR-synthesized images.
• This approach offers a significant advancement in real-time blind quality assessment for FVV.
• The method provides a more faithful representation of human perception for VR and AR applications.