No-reference Stereoscopic Image Quality Assessment Based On Sparse Representation

With the characteristics realistic feeling of scene and media interactive features, stereoscopic video attracts more attentions greatly to meet the demand of our life, and has broad space for development. No-reference stereoscopic image quality assessment is one of the key technologies in stereoscopic video system, and has important research and pratical value. In this paper, the images of the sparse representation model and no reference stereoscopic image quality assessment have been unified, to establish no reference stereoscopic image quality assessment model. The works consist of three parts as briefly summarized below:The first part summarizes the current objective quality assessment algorithms, and focuses on the human visual system in order to study the influence mechanism of perceptual features of stereoscopic image.Secondly, two no-reference stereoscopic image assessment methods are proposed according to the analysis of the characteristics of stereoscopic image: 1) A no-reference image quality assessment method is proposed based on the Bidimensional Empirical Mode Decomposition(BEMD). In the method, Intrinsic Mode Function(IMF) components and residual image are firstly produced by applying BEMD. Then, the statistical information(mean, variance and entropy), and shape and scale parameters are obtained by applying generalized Gaussian distribution method on the intrinsic mode function components, to form stereoscopic image feature information. Finally, Support Vector Regression(SVR) is performed to predict the objective scores by establishing the relationship between the stereoscopic image features and the subjective scores. 2) A no-reference stereoscopic image quality assessment(NR-SIQA) method is proposed based on binocular energy response. In the method, firstly, local features are obtained in different directions, frequencies and disparities by applying Gabor filtering for the distorted stereoscopic images. Then, these features are fused as stereoscopic features. Finally, SVR is performed to predict the objective scores by establishing the relationship between the stereoscopic features and the subjective scores.Last, to establish no reference stereoscopic image quality assessment model with human visual characteristics by depth studying of various types of non-reference image quality assessment methods and the images sparse representation theory, two no-reference image quality assessment methods are proposed based on the sparse representation: 1) A noreference image quality assessment method is proposed based on the Gabor codebook. The method is divided into two processes. In the training process, intrinsic mode function components are firstly produced by applying BEMD on the training images. Then, the image visual codebook is constructed by using feature clustering algorithm. Finally, full-reference quality assessment method is performed to construct the visual quality table of each visual codebook. On the testing process, through calculating the minimum Euclidean distance between each testing feature blocks and the image visual codebook, and weighting the minimum Euclidean distance of visual quality value to obtain the objective evaluation value of the stereoscopic image. 2) A no-reference image quality assessment method is proposed based on the improved sparse representation. On the basis of preliminary studies, the method is divided into two parts.(1) The image visual codebook is constructed by using local area.(2) The image visual codebook is constructed by using global area. On the testing process, through calculating the minimum Euclidean distance between each testing feature blocks and the local, the global image visual codebook, and weighting the minimum Euclidean distance of visual quality value to obtain the objective evaluation value of the stereoscopic image. Experimental results show that, the no-reference image quality assessment method based on the improved sparse representation has good effect on the symmetric and asymmetric image databases, and not need complex machine learning process. It has lower complexity and more accurately in predicting the stereoscopic images quality.