No-reference Image Quality Evaluation For Screen Content Images

With the rapid development of the Internet and the popularity of various electronic devices,people not only have access to traditional natural images in daily life,but also have various screen content images.Natural images are often captured by digital cameras,while screen content images are generated by electronic devices such as computers.Screen content images are inevitably interfered with by various types of image distortion during transmission,reception,and encoding,thereby affecting human visual perception of screen content images.This paper mainly considers the characteristics of the human visual system and the structural characteristics of the screen content image.The research on the image quality evaluation method of the screen content image is carried out,and the three no-reference image quality evaluation methods are proposed on the recognized screen content image database.A good experimental result was obtained.The main results and innovations of this paper are as follows:1.Results 1 proposes a method for evaluating the quality of screen content image based on texture feature and sparse representation.The proposed method is mainly based on three gradient maps of the screen content image: absolute gradient magnitude map,relative gradient magnitude map,and relative gradient direction map.Neuroscience literature has shown that image textures can be obtained by obtaining higher order derivatives,so the texture information of the screen content image can be represented by the local gradient histogram features extracted from the gradient map described above.In the quality score prediction phase,methods of dictionary learning and sparse representation are used.2.Results 2 proposes a method for visual quality evaluation of screen content image based on visual edge model and Adaboosting back-propagation neural network.The method adopts the Gaussian difference model of human eye to obtain the edge information reflecting the visual quality of the screen content image.Through the human visual model,the method calculates edge maps of two types of scales.Then,the L-moment estimation is used to extract the edge features of the image without the reference image.In the final the prediction phase of quality score,Adaboosting back-propagation neural networks with better predictive performance indicators were used.3.Results 3 proposes a method for image quality evaluation of screen content image using stacked auto-encoders in the textual region and the pictorial region,respectively.The pictorial and textual regions are first segmented by the screen content image segmentation method.Second,extract the corresponding quality-aware features from these two regions.Third,we train two parallel stacked auto-encoder networks in an unsupervised manner,and then input these extracted features into the trained stacked auto-encoders to calculate the corresponding deep features.Then,the two support vector regression machines are trained by utilizing the deep features of the two regions and the subjective evaluation scores of the distorted screen content images.Finally,the textual quality score and the pictorial quality score obtained by the two support vector regression machines are combined into a final overall visual quality score by weighting the sum.