Study On Screen Content Image Quality Assessment

With the development of Internet technology,screen content image(SCIs)has been widely seen in people's life,and the quality evaluation of SCIs has become an increasingly important issue.In the past decades,a large number of natural image evaluation algorithms have been proposed and achieved significant achievements.However,research on screen content image quality evaluation is very few,and the quality problem for the detection of SCIs and enhance the user experience of remote computing is important,so we must study the effective quality evaluation method of SCIs.In this thesis,the quality evaluation of screen content image is studied.We propose two methods: naturalization-based screen content image quality assessment and screen content images quality assessment based on edge model and gradient similarity.The main contributions of this thesis are as follows:1.Diff erent from natural scene images,SCI is usually a mixture of text and picture,these are displayed on the screen with some unnatural features,such as thinner profile lines and sharp edges.However,traditional images quality metrics mainly are designed for natural scene,which don't fit well into the SCIs.Motivated by this,we present a simple and eff ective method to naturalize SCIs,which aims to unify SCIs with natural scene images as far as possible on edge features.So that the traditional quality models can be applied for SCI quality prediction.First,the screen content image is pre-processed with the naturalization method,and then the pre-processed image is tested with the existing natural image quality evaluation algorithm.Extensive experiments and comparisons demonstrate the eff ectiveness of the proposed algorithm.2.SCIs include graphics and a large number of text,charts,etc,Compared with natural scene images,screen content images usually have higher contrast and thinner edges.Human visual system is highly sensitive to image edges.With the above considerations,a full reference image quality evaluation algorithm is proposed based on edge model and gradient similarity.The edge model is based on the screen content image segmentation,which extracts two significant features: the edge width and edge contrast.Screen image segmentation is based on difference of information entropy of text and picture.Then the gradient similarity of the image is calculated as the third feature.At last,the edge width,edge contrast and gradient similarity features and the corresponding subjective scores were used to predict the image quality score based on SVR training.The performance of the proposed method is evaluated based on public databases,and the experimental results demonstrate the advantages of the proposed method.