Research On Visual Quality Evaluation Of Screen Image

With the rapid development of Internet technology and cloud computing,more and more screen content images appear over Internet.At the same time,the development of intelligent terminal has accelerated the production of screen content images.Similar with natural images,in the process of screen content image acquisition,transmission and storage,some information may be lost and different kinds of distortion might be introduced in screen content images.In order to obtain a good quality of experience,objective evaluation algorithms should be designed to evaluate the visual quality of the screen content image,and then the performance of various screen content image processing systems can be monitored and optimized.Compared with the research on visual quality assessment of natural images where a lot of related algorithms have been proposed,screen content image quality assessment has just begun.The visual characteristics of screen content images are different from those of natural images,and the visual quality assessment algorithm designed for natural images is not suitable for that of screen content images.Therefore,it is highly desired to design an effective visual quality evaluation method for screen content images.In this paper,we explore the visual quality assessment of screen content images systematically,aiming to propose good visual quality assessment methods of screen content images by perceptual characteristics of the human visual system.The main content includes visual feature extraction,quality evaluation model design,and comparison experiments for performance evaluation.Considering the high cost and non-real time of subjective quality assessment for screen content images,we design a full-reference quality assessment method for screen content images by different characteristics of screen content images from these of natural images.Fullreference image quality assessment aims to compute the similarity between the reference and distorted images.The more similar information quantity or the higher the feature similarity,the better the visual quality of the distorted image is and the better performance the proposed method is.In view of the sensitivity of human visual system to structural information,this work analyzes the visual characteristics of screen content images and calculates the quality map by the similarity between the original image and distorted image based on gradient domain.Furthermore,with the consideration of different perception for different regions in screen content image,a weighting strategy is designed for weighting map prediction by highorder derivative variation.The final quality score of screen content image is predicted by combining the quality map and weighting map.Experimental results on public benchmark show that the proposed full-reference quality assessment method can obtain better performance than other related full-reference quality assessment methods.In most multimedia processing systems,full-reference image information is hard to obtain or there is no any full-reference image information.In this case,an accurate and efficient no-reference screen content image quality assessment method is much desired.Here,we build a no-reference quality assessment method for screen content images by considering the structure and orientation information.Since perceptual features of screen content images do not obey Gaussian or Gaussian-like distributions,we use histogram to represent image structure and orientation features.Finally,we use the machine learning technique to train the proposed model from feature vector to subjective ratings.Experimental results show that the proposed no-reference quality assessment method can predict the visual quality of screen content images more accurately than other related methods.Additionally,with the wide use of deep learning techniques in multimedia processing applications,we propose a no-reference quality assessment method for screen content images based on dual-channel deep neural network.Since the data of screen content images is limited,previous studies of quality assessment for screen content images rarely use deep learning techniques.To increase the training data,we construct image pairs of screen content images to expand the number of training sets and reduce the impact on the training.Then,an unsupervised learning framework is designed to extract perceptual features of the images by a convolutional neural network.Finally,the visual quality score of the distorted image is obtained by the trained deep neural network.Experimental results show that the proposed deep learning based no-reference quality assessment method of screen content images can obtain better performance than other traditional no-reference quality assessment method.