A Study Of Image Quality Assessment Based On Visual Attention And Natural Scene Statistics

With the rapid development of information technology, image is playing a more and more important role in information acquisition, it has a direct relationship with the adequacy and accuracy of information obtained. However, in the process of image acquisition, processing, transmission and storage, degradation problems will been inevitably produced because of the influence of various factors. This brings great difficulty to access information or image post-processing. Then creating an efficient system to evaluate image quality is of great significance. Because Human Visual System(HVS) is the final receiver of image, it is the most reliable way to assess the quality of images, so it will effectively improve the accuracy of objective eva luation algorithm if the effect of HVS has been taken into account in quality evaluation process.At the same time, since Natural Scene Statistics(NSS) is the essential attribute of image, image quality is closely related to its change, This article is mainly based on HVS and NSS theory, the problems how to evaluate an image quality under different application conditions,such as in the absence of reference image or with the reference image,has been researched. An overview of the main research contents and achievements of the thesis are as follows:1. Based on the Discrete Cosine Transform(DCT) and sparse representation, an image quality evaluation method without reference image has been proposed, it mainly solve the problem that the result of image quality evaluation will been inaccurate without the reference image. This method takes the natural degree of image as an measure to assessment image quality, a lot of experiments show that the distribution of DCT coefficients is an important NSS characteristics, it will be changed with image’s distortion,while it also been proved that variation degree of the distribution was closely related with distortion intensity and distortion types. Here we use the fitting parameters of coefficient distribution in DCT trans form domain as features, combined with the method of sparse representation to obtain the image quality. Experiments on LIVE2 prove that, the algorithm can evaluate various types of distorted images without reference images, and the results are highly consistent with subjective evaluation.2. Based on edge information,a full reference image quality assessment method is proposed. Most traditional evaluation methods do not consider HVS characteristics. Through physiological experiments we know that our human eyes pay more attention to edge information when observe image. Based on this characteristic, we improve the traditional classical method Structural SIMilarity(SSIM) by using this visual characteristics. the algorithm proposed here firstly extract the edge information(including Edge gradient. Edge direction. Edge constant etc.) from both the reference image and distorted image,In addition, we combined the image’s local gray information and contrast information to obtain the final image quality. Experiments on LIVE2 show that, in addition to white noise images, this method has a higher prediction consistency than most other full reference algorithms with very low-time consumption.3. Based on the visual saliency, a full reference image quality assessment method is proposed.According to the HVS study, it is found that the responses of human visual system are different when image’s distortion region or distortion degree is different, so according to this characteristics, visual saliency is used as weight in image evaluation process.In our paper, the structural similarity of image is combined with visual saliency, while PSNR measure is also combined with visual saliency. At last, according to the importance of two measures, the final image quality is computed by combining them.Experiments on LIVE2 and TID2008 prove that, this method has obvious advantages in complexity, robustness and consistency with subjective evaluation.