| With the rapid development of multimedia information technology,stereoscopic image/video technology brings a new way of entertainment has become an indispensable part of our daily life.However,in practical applications,it is difficult to obtain the information of the original image..Therefore,it is an important subject to study the no reference stereoscopic image quality assessment(NR-SIQA)model.The thesis starts with the perception factors and establishes three kinds of effective no-reference stereoscopic image quality assessment model by fusing the single and binocular information of image.Finally,in order to improve the generalization ability of the model,an unsupervised stereoscopic image quality assessment algorithm is established.The main research contents are as follows:(1)Aiming at the problem that the existing stereoscopic image quality evaluation models do not fully consider the binocular perception,and considering the low complexity of spatial natural scene statistics(NSS),a generalized no-reference stereoscopic image quality evaluation model based on spatial NSS is proposed.First of all,an improved cyclopean map model is established to fit the human visual system.Then,the objective quality of stereoscopic image is predicted by extracting the spatial NSS characteristics of the cyclopean map and the disparity map.When tested in PhaseI database,the accuracy and monotonicity are greater than 0.94.Experimental results show that the proposed method is more consistent with human subjective perception.(2)Aiming at the problem that the monocular and binocular features affecting the quality of the stereoscopic image,this paper extracts the gradient magnitude(GM)feature and laplacian of gaussian(LOG)feature of the image,and a no-reference stereoscopic image quality assessment model based on the gradient magnitude is established by combining binocular features;Then,considering that the human visual cortical neurons have a high sensitivity to the direction information of the image,and that the gradient direction conveys the complementary information carried by the gradient magnitude,we extract three features from the stereoscopic image,including the gradient magnitude,the relative gradient orientation(RO)and relative gradient magnitude(RM).Then,the no reference stereoscopic image quality assessment model based on gradient direction is established.Experimental results show that the accuracy and monotonicity of the two models are both greater than 0.94,indicating that the proposed two kinds of objective models can accurately predict the objective quality of stereoscopic image.(3)Aiming at the problem that the penalty factor and the radial basis function parameters of the support vector regression(SVR)have great randomness,an objective stereoscopic image quality assessment method is proposed based on the genetic algorithm optimization SVR(GA-SVR)model.Firstly,the single and binocular features of the stereoscopic image are extracted.Secondly,the GA-SVR model is established and we pick out the optimal parameters of the model.Finally,the optimum parameter composition is introduced into the SVR to predict the objective scores of the stereoscopic image.Compared with the image quality of directly with the SVR model,this model has a better performance in experiment.(4)Aiming at the problem that the above-mentioned models have weak generalization ability,an unsupervised no reference stereoscopic image quality assessment method which does not need DMOS is proposed.Firstly,the local features are extracted from the region of interest image blocks and the multivariate gaussian(MVG)model is established.Secondly,the MVG model of left view and right view are established by the similar method.Finally,by calculating the distance between the measured image and the natural image,the objective quality is evaluated by gain-control model.The experimental results show that the evaluation performance of this model has been improved remarkably compared with other unsupervised models. |
PDF Full Text Request