Research On No-reference Image Quality Assessment

As an important carrier of information,the image has been widely used in health and medical community,consumer electronics,etc.However,it is inevitable to induce distortions during image acquisition,transmission,processing and display,causing quality degradation.How to effectively evaluate,compare and optimize the image quality has gradually become a research hotspot in many fields,such as visual psychology,image processing,pattern recognition and artificial intelligence.In this thesis,we investigate the issue of quality assessment of 2D images and stereoscopic 3D(S3D)images.The main contributions of this thesis are listed as follows:In the aspect of 2D image quality assessment(IQA),we mainly focus on the issues about distortions of Gaussian blurriness and tone mapping.By analyzing the relationship between the gray level of a central point and that of its adjacent points,we first code the whole image and extract the statistical feature as well as entropy information of the local binary pattern map.Then,a no-reference(NR)blurriness assessment model is built to fuse the feature vector againest its subjective score via the support vector regression(SVR).Experimental results demonstrate that,the proposed method is fast and can accurately predict the subjective score.Also,it is superior to all competing methods.By analyzing the preservation ability of the distorted image on color,texture,structure and naturalness,we extract multiple quality-sensitive features and propose two NR quality assessment methods by fusing the extracted features via machine learning.One is inspired by the simulation of color information processing mechanism in visual channel,and another is built by considering multiple image properties,i.e.,colorfulness,structure and naturalness.Experimental results demonstrate that,both methods correlate well with the subjective scores and are superior to the mainstream competing methods.In the aspect of S3 D IQA,we mainly focus on the quality assessment issues of S3 D images and DIBR-synthesized images as well as the visual discomfort issue caused by S3 D subtitle.As for the S3 D image,we first generate the cyclopean view.Then,we quantify the naturalness of the left,right and cyclopean views.Meanwhile,the similarity and difference between the left and right views are quantified to indicate the degree of the asymmetric distortion.Finally,a NR quality assessment method is proposed based on SVR.Experimental results demonstrate that,the proposed method can accurately predict the subjective score.Meanwhile,it can effectively evaluate both the symmetrical distortion and asymmetrical distortion,and is superior to the mainstream competing methods.As for the DIBR-synthesized image,we propose a quality assessment method by evaluating three distortion types,i.e.,the disoccluded region,boundary stretching and global sharpness.By combining the above three measures linearly,the proposed method can address the quality assessment problem of DIBR-synthesized images efficiently and effectively,and outperform the mainstream competing methods.As for visual comfort issue caused by S3 D subtitle,we first investigate S3 D subtitle position's influence on the visual comfort based on subjective experiments.Then,accrording to the subjective experimental results,a novel region selection method for the S3 D subtitle is proposed.Experimental results demonstrate that,the proposed method can effectively reduce the visual discomfort caused by S3 D subtitle without affecting the video content.