Research On Stereo Image Quality Assessment Based On Sparse Representation

With the development of stereo imaging technology,stereo images are gradually coming into our lives.However,distorted stereo images can cause psychological and physiological discomfort,which severely limits the development of stereo imaging technology.Therefore,finding a systematic and effective algorithm to evaluate the quality of stereo images has become a research hotspot in related fields.The main work of this paper is as follows.First,in this paper,we proposed an effective 3D image quality assessment method based on an adaptive cyclopean image by using ensemble learning.Considering the gain control and gain enhancement human vision mechanism,the left and right views of stereo image are fused into a cyclopean image to simulate the fusion process of the binocular vision information in visual pathway.Our cyclopean image takes into account the linear and non-linear fusion process of left and right views,and it is suitable for symmetrically and asymmetrically distorted stereo images.Our cyclopean is an adaptive fusion image model.In addition,considering the visual saliency mechanism,the salient map is used to modify our cyclopean image to let the salient area of our cyclopean get more attractive.As a result,we can get better results.To remove redundant information out from our cyclopean,the sparse representation is applied to extract essential features of the cyclopean image.At last,in order to get better regression accuracy on extracted sparse features,the ensemble learning is used to get the final quality score of stereoscopic image.The ensemble learner can improve the regression accuracy by 2% than a single learner.Second,in this paper,a sparse binocular fusion convolution neural network is proposed to evaluate the quality of stereo image.The network simulates the complex fusion processing process and quality perception process of binocular visual information from retina to visual cortex.The left and right views fuse four times in the network to simulate the long-term and complex fusion process of binocular visual information in human brain.The multi-convolution layer and multi-branch of the network simulate the layered and parallel processing of visual information in human brain.And the final full connectivity layers of the network simulate the process of quality perception of the fused binocular visual information.In addition,in order to overcome the computational-intensive and memory-intensive problems of convolution neural networks,a structural sparsity learning(SSL)method is used to regularize each convolution layer of proposed convolution neural network.SSL can remove the redundant weights in the network,speed up the network and improve the performance of network.It can achieve 2.0× speedups on LIVE I database and 2.3× speedup on LIVE II database on the basis of improved performance.In order to verify the performance of the above proposed quality evaluation algorithms,we carried out experiments on the LIVE 3D Phase I,LIVE 3D Phase II,Waterloo IVC 3D Phase I and Waterloo IVC 3D Phase II databases.Experimental results show that our proposed algorithms can simulate the human visual characteristics well,and the evaluation results on symmetrically and asymmetrically distorted stereo images are excellent.