Metric Method, Based On The Hvs Several Image Quality

Images have become the most important source for people to acquire information from outside world because of its certainty, visualization, high-efficiency and wide-adaptation. A correct evaluation of image quality is a very significant research in image information processing. The subjective methods for image quality are consistent with actual instance, but cost too much time, and the experiment results often change with the experiment condition. The existing objective evaluation methods for image quality, for instance, PSNR and MSE, avoid defects which the subjective methods have, and possess simple merits which are easy to implement, but they are not exactly in consistent with human's visual characteristic. So any reasonable objective quality evaluation method of image should abide fully by human vision system. In this paper, based on the former subjective and objective methods, I study this issue and gain some results as follows:1. A novel image quality evaluation method is presented, which is based on the local self-similarity, a more essential and inner feature of images, and the fact that the attraction on images decreases from center to margin when people are browsing through them. We choose collage error as local self-similarity evaluation of the image, because collage error visually represents the similarity between value range blocks and optimal matching blocks. Thus, the differences between original image and distorted image appear as the differences of their self-similarity. The weighted value is connected with the position of value range block in the image and the grey changing level of value range block. The way we chose is consistent with HVS characteristic. While people browsing through pictures, the attraction is attracted first by the central part then spread to margin. The importance of the image is decreasing from center to margin, this is consistent with what we chose while we take pictures (always set the concerned object or part to the center).We numerically describe the consistency between CEM and MOS using correlation coefficient. We provide PSNR and FI and even calculate the correlation coefficients between them and MOS using the evaluation values for images to be valued. Considering the differences between the types and the scales of noises, we conclude the correlation coefficients of the threegroups. The correlation coefficient of CEM and MOS for stripe images, block images and decoding images is 0.97, 0.91, 0.95. And the corresponding coefficient of PSNR and MOS is 0.91,0.90, 0.90. But FI cannot judge stripe images and block images, and the coefficient of FI and MOS is 0.92. Obviously, CEM is much better than PSNR and FI. Experiment results indicate that the proposed method provides a more effective judgment for different types of distorted images and a better consistency with human's visual characteristic. 2. Make use of hypothesis testing technique, we propose linearity regression model of image quality measurement, at the same time, we discuss synthesis image quality method of five errors which image distortion is affected by mainly. The five errors are luminance coding error, spatial frequency weighting error, random error and disturbance, correlated error and local error. This method combines subjective method with objective method. We do regression analysis by a lot of experiment results. At last, we gain simple and feasible expression, which is be linear with spatial frequency weighting error, correlated error and local error. The method is an effective and reasonable image quality measurement.