| With the development of multimedia application technology,the visual content information received by consumers on mobile terminals has become more and more abundant,which constitute a new type of image called screen content images(SCIs),such as remote desktops,online meetings,etc.However,because of the imperfection of wireless communication technology,these images will inevitably introduce a variety of interferences during the propagation process,which will destroy consumers’ visual experience.Because the quality score prediction of the SCIs conduces to supply a reliable basis for the user to present high-quality images,it is very necessary to design a SCIs quality assessment model that conforms to the human visual system.Taking into account the characteristics of SCIs compared with other types of images,this paper proposes two different quality assessment models for SCIs,as follows:The first method is a full-reference quality assessment model for SCIs based on the concept of global-guidance and local-adjustment.For the first time,this method uses a fully convolutional neural network to segment the SCIs into textual and pictorial region.For the former,this method firstly proposes the concepts of edge expansion and edge step to describe the structural characteristics of the text,and the quality score of the textual region is measured accordingly.For the latter,this method proposes to use the just noticeable difference(JND)to describe the perceptual characteristics,and judge the quality score of the pictorial region based on this.Finally,this method innovatively designs a score integration method based on the concept of global-guidance and localadjustment strategies to better integrate the textual and pictorial quality scores to generate the final quality score.The second method is a no-reference quality assessment model for SCIs based on the stage-learning.In the first stage,based on multi-task learning,the model obtains the ability to recognize image distortion information and uses it as a priori knowledge for image quality evaluation.In the second stage,the knowledge learned by the pre-training network is transferred to the image quality assessment task,so as to solve the problem that the number of samples in the current SCIs database is not enough to train the neural network.In order to support the first stage of training,this paper also established a larger-scale distorted SCIs database.Finally,this paper conducts sufficient experiments on the mainstream database of SCIs.The experimental results show that these two methods have achieved substantial performance improvements and are more suitable for SCIs quality evaluation. |
