Compressed Distortion Oriented Visual Quality Assessment In Mixed Scenes

With the rapid development of computer,network communication and multimedia technology,various multimedia applications are integrated into users' daily life and work,such as desktop sharing,online conference,live game broadcasting,online shopping,virtual roaming,etc.Meet the user's application needs,visual content extends from traditional natural scene to mixed scene,which is composed of computer-generated content,natural scene content and textual content.At the same time,the field of view extends from fixed range to panoramic range to achieve immersive experience.To fulfill the requirements of real-time interactive applications,mixed scene content needs to be efficiently coded and compressed for reducing bandwidth and user costs.However,the visual quality distortion introduced by the compression of mixed scene content can significantly affect user experience.The research on compressed distortion oriented visual quality assessment in mixed scenes can promote the popularization and development of this kind of application.In this paper,we investigated the image quality assessment problem on compression artifacts,including the quality evaluation of screen content images related to desktop sharing,and the exploration of the quality evaluation of panoramic images related to virtual roaming.These contents all present the characteristics of mixed scenes,which include the natural scenes captured by camera and the scenes generated by the computer.At present,there are various of image quality assessment algorithms,but these algorithms are not suitable for the visual quality assessment in mixed scene.Therefore,this paper carries out research on these problems.The main contributions are summarized as follows:(1)By analyzing screen content image and natural scene image quantitatively and qualitatively,it is found that screen content image contains natural scene content,textual content,computer-generated content,and the sensitivity of human eyes to different contents is different.In addition,screen content image does not follow the statistical characteristics of natural scene image.According to these characteristics,this paper proposes a reducedreference quality assessment algorithm based on visual sensitivity combined with the statistical characteristics of screen content images.Experiment results show that the performance of the proposed reduced reference quality assessment algorithm is better than that all of the existing algorithms.And the algorithm can be compared with most full reference quality assessment algorithms.(2)To investigate the influence of screen content video on the visual attention and quality preference,we build an eye fixation dataset for screen content video.Besides,we proposed a no reference video quality assessment algorithm for screen content.Experimental results show that the performance of the proposed algorithm is better than existing no reference algorithms.(3)With the popularity of virtual reality,the stereoscopic display method,which is different from the traditional two-dimensional plane,brings new challenges to quality assessment.In this wraparound 360o×180o display situation,the user's viewing behavior pattern is more complicated,and the factors affecting quality assessment are more complicated.In this paper,the quality evaluation of panoramic images under compression distortion is explored,and the first database for Just Noticeable Difference(JND)is established.Based on existing models,a prediction algorithm for JND of compressed panoramic image is proposed,and it is tried to be used for the quality assessment of panoramic images.The increasing popularity of visual content in mixed scenes have made the related image technology a research hotspot.However,most of the existing quality evaluation algorithms are aimed at natural scene content,which is dwarfed when used to evaluate visual quality in mixed scene.In this paper,a lot of work has done to analyze the characteristics of these visual contents and the behavioral characteristics of human vision during evaluating these contents.Reducedreference and no-reference quality assessment algorithms based on visual characteristics are proposed,and an improved JND model is proposed to explore quality evaluation of panoramic images.