No Reference Image Quality Assessment With Multi-kernel Learning

In recent years, with the rapid development of hardware resources, people’s daily information is increasing rocketly. The internet is filling with all kinds of images witch is transported as visual information carrier. During all kinds of image processing methods such as,acquisition, compression, transmission often introduce distortion to images, so the image quality evaluation is an important index of the image and video processing system, video equipment performance testing. Also, image quality evaluation can also be used for intelligent monitoring, automatic focus to provide effective reference. Studying with an image evaluation method witch is consistent with human visual perception has a good application prospect.SVM(Support Vector Machine, SVM) promote the rapid development of the kernel methods successfully, it has gradually penetrated the many areas of the Machine learning. In response to a higher dimension, a more complex data, method of multi-core also become a research hotspot. In this paper, based on the multi-core learning method, a no reference image evaluation algorithm witch faces to multiple kinds distortion type of image distortion is studied, the main work is as follows:1. Blind Image Quality Assessment with Texture Features Via Multiple Kernel Learning. This method put forward the concept of structure tensor flat areas, and add to the range of texture feature extraction, combined with phase, using the modified Gray-gradient Co-occurrence matrix and Gray-level Co-occurrence Matrix, we extract multiple kinds of texture characteristics, meanwhile it is difficult to match the specific texture characteristics with the distortion through the effective combination experiment by single-kernel machine learning method, and single-kernel learning methods can’t modeling effectively on the texture characteristics, so we choose a learning machine based on layered multi-core to training and studying, and we test it on various image library through cross-validation,and we can also get a stable evaluation results. And we use the flat image quality evaluation algorithm as mentioned above, and describe the characteristic curve of distortion in the image when the second distortion joins in, also we analyze the situation that two kinds of distortions exist at the same time, and the variational characteristic curve of image shows the feature extraction algorithm can seize the distortion in the abnormal part of natural images.2. Stereo Image Quality Evaluation With Texture Features via Multi-core learning. Firstly we describe the disparity map extraction method, and then according to the principle of the method, and add the error energy diagram. Secondly we detail the statistical characteristic of error energy curve, and we find that it has good correlations with the actual situation, showing that new features have well effectiveness. And we use the evaluation method of the above to evaluate the disparity information. We conduct random sampling test repeatedly over LIVE stereoscopic image database through the multi-kernel machine learning, and the algorithm also has a well performance.4. Image Quality Evaluation With Nonsubsampled Contourlet Transform via Multi-core learning. This method use the Nonsubsampled Contourlet Transform to extract the subband from multi-channel and multiple directions, and we reduce the decomposing directions, by using the amplitude of Fourier transform, the average gradient, information entropy, and multi-kernel machine learning mentioned above, we have good resuls on the LIVE database.