Research On Objective Image Quality Assessment Methods Based On Local Visual Features

Image information is one of the main routes for human beings to understand the outside world. During the process of image data acquisition, transmission, compression and processing, various distortions will be introduced to result in image degradation, and decrease the visual effect of the image. Image quality assessment method can help to design the parameters of imaging system and to optimize the algorithms, and thus how to evaluate the image quality reasonably is becoming one of the hot topics in the field of image processing. This dissertation focuses on representation and extraction of features of image distortion, and explores to use local visual features to model the distortion information while considering the characteristics of the human visual system. It mainly studies the full-reference image quality assessment method and no-reference blur image quality assessment method and applies the image quality assessment method to medical image fusion. The main work and innovations are as follows:A new image quality assessment method based on local information distortion modeling is proposed in this dissertation as traditional structural similarity image quality metric cannot evaluate noisy image and severely blurred image accurately enough and does not correlate well with human subjective perception. This method separates image distortion into three types, which are pixel grayscale distortion, local contrast distortion and local structure distortion. The gray level instead of neighborhood average of pixel is used to model the grayscale distortion, the local binary pattern is used to model the local contrast distortion and the local variance is employed to model the local structure distortion. This method owns a simple calculation and behaves well on 5 types of distortions in LIVE database, indicating a good evaluation performance. In addition, since the structure information term of SSIM method is too simple, a Riesz transform based structural similarity image quality method is proposed. According to the fact that local structure information can be well represented via Reisz transformation, this method uses Riesz transform to extract the first-order and second-order image features to rebuild the structure information term to obtain the improved structural similarity image quality metric. Compared with SSIM and other methods, this metric has a better evaluation performance.The subjective perception of luminance for human eyes obeys the Weber’s law, which says that the noticeable stimulus change is proportional to the original stimulus. An image quality assessment method based on the human visual characteristics is proposed, in which the differential excitation map of image is used to model the Weber’s law, the visual saliency map is used to model the visual attention strategy, and the impact factors to human judgment with different viewing conditions. Further, since the calculation of differential excitation does not cater the contrast information, the image gradient is used as a complementary feature. A differential excitation similarity based image quality assessment metric is proposed based on the two features. At last, a log-Gabor Weber features is presented to better extract the local information of image, and it is used to design an image quality assessment method that can evaluate both the color and grayscale images. The proposed algorithms above have very good performance on several benchmark image databases such as LIVE and CSIQ.A blur image quality metric based on local standard deviation and saliency map is proposed due to the high computational complexity of existing no-reference blur image quality assessment method. Firstly, the reblur effect of image is used to construct a reference image with a Gaussian low-pass filter. Then the two simple features namely the local standard deviation map and saliency map are used to evaluate the blur image quality according to how much they change after reblurring. The experimental results show that the proposed algorithm correlates well with human subjective perception and it owns a close evaluation performance and much lower computational complexity compared with the state-of-the-art no-reference blur image quality assessment method.Fusion of multi-modality medical image is a hot topic. As existing image fusion quality assessment indices are not well consistent with human subjective perception, a medical image fusion quality assessment method based on monogenic features is proposed. The proposed algorithm takes advantage of the monogenic signal which can well represent the local information of image to produce a new kind of phase congruency namely monogenic phase congruency. Then a monogenic feature similarity based image quality assessment method is presented which takes into account the salient edge information and monogenic phase congruency information. The monogenic phase congruency map of original image is used as weighting factor to simulate the different importance of different regions for human eyes, so as to get the weighted monogenic feature similarity. Finally, entropy and edge energy of the original image are used as the importance factor to obtain the quality evaluation index of fused image. Compared with other methods, the proposed method is more consistent with the subjective human visual perception.