Research On No-reference Stereo Image Quality Assessment Based On Visual Perception

With the rapid development of stereoscopic images,stereo products including stereoscopic movies and games and so on greatly enrich our life experience.However,stereoscopic images can suffer from quality degradations during the acquisition and transmission,making viewers feel uncomfortable.Therefore,the study of stereoscopic images quality assessment has great significance.In this paper,we have designed an objective quality evaluation algorithm of stereo image via combined features based on human visual perception.First,the stereo image pair is decomposed by wavelet-packet which has accurate resolution,afterwards the decomposed left and right views are fused to obtain cyclopean map and difference map based on the principle of binocular rivalry and binocular suppression.Then,Natural Scene Statistics features and information entropy are extracted on the fusion map;besides,the structural similarity feature is extracted by taking into account the internal relations between the left and right views.Finally,the Support Vector Regression is used to establish model between the perception features and the subjective scores,which can predict the objective evaluation score.The experimental results on LIVE 3D image databases show that the proposed algorithm has high consistency with the subjective evaluation results,it outperforms state-of-the-art stereoscopic image quality assessment algorithms.Meanwhile,we have proposed a stereo image quality assessment model based on deep convolution neural network.First,the left and right views of the stereo image are divided into patchs,and depth convolution neural network is employed to extract features of the left and right viewpoints.Second,the features are fused and deliver to the full connection layer regression network to generate each patch's score and its corresponding weight.Finally,the final stereo image quality is obtained by multiplying the patch quality and the corresponding weight.Experimental results show that the proposed algorithm model is consistent with the subjective evaluation of human,and is in accordance with the human visual perception characteristics.