Gradient Structure Similarity Based On Visual Perception Image Quality Evaluation

Image quality is an important basis for a measure of image processing algorithms good or bad and optimize system parameters, thus building an effective image quality evaluation methods in the field of image acquisition, encoding and transmission is very significance. In recent years, with the rapid development of image processing technology, image quality assessment received extensive attention of scholars both at home and abroad and corporate, sprung up many image quality assessment algorithm, the quality of the typical model are based on human visual system evaluation model and quality evaluation model based on structure similarity, etc. Current joint video team has the quality assessment algorithm based on structure similarity as the important indexes for evaluation of image quality, introduced to a new generation of video compression coding standard h.264, the calibration model.Image quality assessment is divided into two kinds of subjective and objective evaluation methods. Among them, the subjective evaluation method need more repeated experiments, the time is long, cost is high, the operability is poor. And objective evaluation method has low cost, simple operation and easy to embedded implementation etc, and become the focus of the researches on the current image quality evaluation. In recent years, based on human visual system model, puts forward the related image quality evaluation method. Structure similarity method mainly based on the independence of the relationship between light and object structure, the brightness and contrast separates from the image structure information, combined with the structure of the image information to evaluate the image quality. This method to some extent, avoid the complexity and multi-channel image content and other issues, trying to simulate the human visual system in extracting the object structure shown by the visual capabilities and features as a whole, and the algorithm is relatively simple, easy to implement embedded.Structural similarity algorithm although objectively reflect the changes in image structure information, but there is still lack of the following:first, the algorithm only focus on the structure of the image information, while ignoring the underlying visual characteristics of human visual system, resulting in its evaluation of the quality sometimes differential subjective scores is inconsistent; Second, the modeling process of the algorithm is simple, no high-level visual simulation of the human visual system characteristics, thus in serious distortion of the evaluation of blurred image, it is difficult to get satisfactory effect.To solve these problems, on the basis of structural similarity assessment model, we propose a structural similarity image quality assessment model based on visual perception gradient (MG_SSIM). The main research work and features are as follows:1) SSIM algorithm focuses on the changes in the structure of the image information, and does not consider the relevant characteristics of the human visual system, sometimes resulting in less accurate evaluation. So, based on SSIM algorithm, using the image error of visibility and contents of visibility, underlying visual characteristics of HVS model, constructing the visual perception function, making evaluation result consistent with subjective feeling better;2) The SSIM algorithm is through the subordinate relationship between image pixels to characterize the structure of the image information, the evaluation results sometimes do not agree with DMOS. Isotropic position of sobel operator weighted coefficient more accurate, detecting edge in different directions, the isotropic sobel operator gradient amplitude behaved more consistency, characterization of the gradient effect is better. So, in this paper, using the gradient map to redefine structure component, instead of the original contrast components and structure components of SSIM, construct the structural similarity model which based on gradient, make the evaluation results more accurate.3) Using the error of visibility and error of content structure function of visual perception, as weight applied to the similarity function, and using the isotropic sobel operator on the image to redefine the structure of the components, on this basis, construct new structure similarity function MG_SSIM.4) Based on PSNR and SSIM, MG_SSIM three algorithms, do experiment on five types of distortion in the LIVE image library respectively, to test effect of this algorithm MG_SSIM compared with the rest of two algorithms, the experiment results show that, the SSIM MG_SSIM model is more accord with human visual characteristic, can better consistent with the DMOS, evaluate the blurred image of severely damaged accurately than SSIM.