Subjective Quality Assessment Of Stereoscopic Images Based On Bayesian Theory

With the popularization of stereoscopic images and video,the comfort quality of stereoscopic images has attracted more and more researchers' attention.Methods of three-dimensional image comfort quality evaluation can be divided into two categories:subjective evaluation methods and objective evaluation methods.Since the human visual system(HVS)is the final receptor of images,the MOS(Mean Opinion Score)value obtained by subjective evaluation experiments is most accurate and reliable.Besides,results of the subjective experiment can also be used as a criterion to test the validity of the objective evaluation method/model.However,the image comfort score of subjective evaluation experiment may be influenced by the evaluator's perference,the experimental environment and the inherent random of the subjective experiment,causing a variety of errors.As a result,how to use the limited subjective experimental score data to calculate relatively accurate and reliable image comfort score,has became an important research topic.Therefore,this paper mainly focus on how to extract the preference features of evaluator from the stereo images subjective experiment's data and apply them to obtain a more accurate and reliable image subjective comfort quality.For this purpose,this paper first established a three-dimensional images database under geometrical distortion for subjective evaluation experiments.When constructing the stereoscopic image dateset,use the baseline distance between stereo cameras and the included angle of the optical axis as the geometrical parameters,to describe the relative position of the left and right image.The image dataset contains 16 kinds of? and ? geometric distortions images of 12 stereoscopic scenes with different parallaxes and regions of interest distribution.Based on the image dataset,after the screening and training of the evaluators that participated in the subjective experiment,subjective comfort evaluation experimental data of 20 subjective about single stimuli and pairwise comparisons were collected.For pairwise subjective experimental data,the existing transitivity detection algorithm can measure the consistency of the evaluater himself's score data and obtain the transitive satisfaction,which can be criterion of evaluation reliability of the scoring data.However,it is difficult for transitive satisfaction to convey the individual's specific preferences that existing on subjective experiments.In particular,the evaluator's tolerance of certain types of parallax distributions or distortions may be different from ordinary people.Based on the limitations of transitive detection algorithm,this paper presents the evaluation consistency algorithm of paired subjective experimental data,which is used to measure the consistency of evaluation data among all evaluaters.And then the consistency of evaluation data is obtained.In this paper,transitive satisfaction and consistency satisfaction are combined to calculated the reliability weight,which is used to measure the reliability of evaluator's scoring data.The subjective experiment results shows that,the reliability weight can measure the confidence of the experiment evaluators more accurately so that the MOS value obtained from the pairwise comparison subjective experiment result is more close to the true image subjective comfort quality.For single stimulus subjective experimental data,International Telecommunications Union has developed a standard method for calculating the image subjective MOS,in which the outlier detection algorithm only considers the inherent randomness of subjective experiments and ignores the possibility that the evaluators may perform specific personal preference features during the experiment,which includes the information of their rating bias and rating data confidence.In this paper,the relationship between geometric distortions and image subjective quality and the distribution of people's preference distribution are taken as priori knowledge,to establish a rating preference model for evaluators and a subjective comfort quality evaluation model for stereoscopic images in a single stimulus subjective experiment.On this basis,using the maximum posteriori probability method of Bayesian Probability Theory,the parameters of the models are solved.Finally,the subjective comfort quality MOS value of the images and the evaluators' preference characteristic parameters is obtained.The analysis of scatter plots of the evaluators' preference and confidence distribution shows that,the models that established in this paper can correctly represent the personal preference characteristics of the evaluators and can be used to update experimental scoring date.Besides,for those date with large randomness and can not updated by evaluator's perference parameters,should be assumed as outliers.Compared with the ITU method,the proposed models obtained MOS value is closer to the true image subjective comfort value.