Objective Quality Assessment Of Stereoscopic Omnidirectional Image

The maturity and commercial era of 5G technology have greatly promoted the development of virtual reality(VR)technology.Stereoscopic Omnidirectional Images(SOIs),as a new media form with an increasing proportion of VR content,have gradu-ally penetrated into every aspect of daily life,besides,it is a relatively mature technology than augmented reality and mixed reality.As one of the research hotspots in image pro-cessing,the main contradiction lies in how to save bit rate while satisfying user's visual experience.Image degradation will inevitably introduced in the process of acquisition,processing,transmission,storage and display,coupled with the unique characteristics of SOI and the complex distortion superimposed with traditional distortion.Exploring the distortion of a specific image format and effectively measuring it will play an important role in image compression and transmission,and improving user's satisfaction.This thesis explores the SOIs as follows:1.We propose a quality assessment model based on projection invariant feature and visual saliency feature for SOIs.At first,the monocular and binocular features of SOI are derived from the Scale-Invariant Feature Transform feature points to tackle the inconsistency between the projection formats and the viewports.In the next,the visual saliency model,which combines the perceptual factors such as chrominance,contrast and structure similarity etc.,is used to facilitate the predic-tion accuracy.In addition,according to the characteristics of the panoramic image,we generate the weight map and utilize it as a location prior,which can be adapted to different projection formats.In the end,the proposed SOI quality assessment model fuses projection invariant features,visual saliency based feature,location prior.Experimental results on both the NingBo University SOI Database and Stereoscopic OmnidirectionaL Image quality assessment Database demonstrate the proposed metric on equi-rectangular projection format outperforms the state-of-the-art schemes.Meanwhile,the proposed algorithm is extended to another five representative projection formats and achieves superior performance.2.From the data-driven perspective,we use dictionary learning method to extract the learned representations of SOI,and then merge with the traditional handcrafted features to measure the distortion of SOI.In order to further solve the problem of inconsistency between subjective and objective evaluation objects,both the data-driven based feature,i.e.,the dictionary learning based fature and handcrafted feature are performed on the viewport that sampled from the sphere,thereby avoiding additional projection conversion operations and adapting to different pro-jection formats.The performance of the proposed quality assessment model on Stereoscopic OmnidirectionaL Image quality assessment Database is better than the current mainstream onesThis thesis conducted an in-depth study on the objective quality assessment of SOI The innovation can be concluded as follows.For omnidirectional image,the learned representations and perceptual characteristics of omnidirectional image are integrated,visual saliency and user's behavior preference when viewing SOIs have been considered For stereoscopic image,a feature point matching method is proposed to detect the perceptual depth distortion.For the composite distortion of the two image formats,the projection invariant related features are utilized.Finally,in order to make the proposed quality assessment model applicable to various projection formats in reality,an adaptive projection strategy is proposed in both work.