Research On No-reference Stereo Image Quality Assessment Algorithm Based On Machine Learning

With the rapid development of high performance computer and multi-media technology,It's more easy to access 3D digital image and video in people's life, such as in the 3D film,intelligent robot, teleconference, industrial control and so on. It's difficult to avoid distortion in the collection, transmission, process and compression of image. The target of researching image quality assessment(IQA) is to design a system that can predict image quality automatically. The predicted quality of computer should be highly relevant with human subject perception. 3D IQA is different from convenient 2D IQA as there is disparity information and depth information in 3D case. Besides, human begin to study 3D IQA later than 2D case and accumulated less experience.As its good generalization performance, SVM has been adopted to the research of image quality assessment. SVM can make feature mapping implicit by introducing kernel function.Mutiple kernel learning(MKL) can solve high dimension,heterogeneous and complex data better than SVM by introducing different kernel function or different kernel function parameter. Extreme learning machine(ELM) is an improvement of a single hidden layer feedforward neural network with the advantages of easy parameter selection way, running fast and so on. The ELM with kernel can solve unknown map problem better by introducing kernel function. The application of machine learning algorithm in stereo image quality assessment is focused in the paper. The main research achievement is in the following.1. It proposes a no-reference(NR) image quality assessment algorithm based on SimpleMKL and Gabor transform. Firstly, the method obtains evaluated disparity map by adopting the SSIM-dense stereo matching algorithm. Secondly, right view image, left view image and disparity map are processed by Gabor transform. Thirdly, entropy and energy of Gabor wavelet coefficient are computed. At last, these features drive SimpleMKL to predict tested image quality. The experiment indices that the framework of utilizing the 2D information and disparity/depth information can show some superiority.2. It proposes a no-reference image quality assessment algorithm based on SimpleMKL and Spatial domain feature. Firstly, evaluated disparity map, disparity-compensation map and confidence map of stereo pairs are obtained by the SSIM-dense stereo matching algorithm.Secondly, left view image, right view image and confidence map are filtered by gradient magnitude operator and Laplacian of Gaussian operator. Gradient magnitude response and Laplacian of Gaussian response of these maps are obtained. Thirdly, all responses are jointly normalized and marginal probability and dependency marginal probability are computed. At last, these features drive SimpleMKL to predict tested image quality. The experiment shows good relation with subject perception.3. It proposes a no-reference image quality assessment algorithm based on ELM with kernel and quaternion wavelet transform(QWT). Firstly, disparity map,disparity-compensation map and confidence map of stereo pairs are obtained by the SSIM-dense stereo matching algorithm. Secondly, left view image, right view image,disparity map and confidence map are processed by QWT and weighted standard deviationand energy of the QWT coefficient of the third phase are computed. Third, the entropy and median of disparity-compensation map are computed. At last, these features drive ELM with kernel to predict tested image quality. The experiment indices its performance is superior and it captures and utilizes disparity/depth information in a proper way.