No-Reference Stereoscopic Image Quality Assessment Method

With the rapid development of 3D technology,3D movies and TV play an important role in daily life and attract global attention.However,during the acquisition,transmission,and storage of 3D images,different degrees and types of distortion will be introduced,resulting in the degradation of image content quality and affecting the quality of experience.Different from 2D images,there is a certain disparity between the left and right views of the stereo pair,which can provide additional depth information and a immersive environment to viewers.However,excessive disparity may cause dizziness and other uncomfortable feelings to viewers.Therefore,the stereoscopic image quality assessment is more challenging than 2D image quality assessment.This thesis mainly studies the stereoscopic image quality assessment method from the following two aspects.Aiming at the problem that the current no reference stereoscopic image quality assessment algorithms are insufficient in the prediction accuracy,this thesis proposes a no-reference stereoscopic image quality assessment method based on binocular visual characteristics and depth perception.This method first discriminates different binocular behaviors by comparing the structural similarity and the amount of useful information between left and right views,and constructs corresponding cyclopean views to simulate the binocular vision mechanism.Then extract the monocular and binocular features from the stereo image pair in the spatial domain.In addition,this thesis proposes a weighted disparity map and longitudinal correlation coefficient map that can measure the depth quality of stereoscopic images and extracts depth perception features from them.Finally,this thesis uses the adaptive enhancement algorithm based on support vector regression to construct the mapping relationship model from the feature domain to the quality score domain.And this thesis evaluates the performance of the proposed algorithm on four public stereo image databases.Compared with some state-of-the-art full-reference,reduced-reference and no-reference stereoscopic image quality evaluation algorithms,the proposed stereoscopic image quality evaluation algorithm has more accurate prediction performance and higher consistency with human subjective perception.In order to further improve the performance of the stereoscopic image quality assessment model,a blind stereoscopic image quality assessment model which extracts quality-aware features in spatial domain and transform domain is proposed.This model extracts the natural scene statistical features in the spatial domain and transform domain at the same time.In addition,considering that the distortion is not evenly distributed in all areas,in order to more effectively simulate the binocular vision mechanism,this thesis first divides the left and right views into blocks,and then discriminates and constructs the binocular combination model to simulate the fuse process in the brain,and extracts the binocular features in the transform domain.Finally,an adaptive enhancement algorithm based on support vector regression is used to train a stereoscopic image quality evaluation model from the feature domain to the quality score domain.Finally,this thesis compares performance with some state-of-the-art full-reference,reduced-reference and no-reference stereoscopic image quality evaluation algorithms on the four public stereo image databases to verify the superiority performance of the proposed algorithm.