Non-reference Stereoscopic Image Quality Evaluation Based On Sparse Representation Of Content

With the rapid development of computer and multimedia technology,three-dimensional digital images have gradually entered people's lives.It's necessary to design a system,which meets human visual perception,to automatically predict stereoscopic image quality owing to the unavoidable distortion caused by the acquisition,transmission,and compression of stereoscopic images and the complexity and high cost of the subjective quality assessment methods.In view of the situation,the paper proposes a non-reference stereoscopic image quality evaluation algorithm based on sparse representation of content by combining the knowledge of human vision system,sparse representation with image area division.First of all,the paper proposes an over-complete dictionary design method based on the self-encoder working principle to overcome the shortcomings of the traditional redundant dictionary construction method in obtaining fixed dictionary.The steps are as follows:firstly,choosing 10 images with large structural differences to chunk;secondly,using the self-encoder framework to extract the most basic features of each block of image;finally,learning the most basic features we obtained and clipping them based on the strategy that it does not have much effect when the energy of image block is small.Experiments show that the over-complete dictionary designed in this paper is more concise and comprehensive compared with the previous dictionary.In addition to,the performance of the over-complete dictionary is better than others.Secondly,the paper proposes a non-reference image region partition algorithm based on convolution neural network after analyzing the previous image region demarcation and edge detection algorithms,which makes up the defect that image region partitioning needs original image.The steps are as follows:firstly,using convolution neural network to detect edges of images at different scales;then,getting a better edge by combining edge over-complete dictionary constructed with image edges of different scales,which is more suitable with human eye perception;finally,using the low threshold value of Canny edge detection and image matching to divide accurately image edge,texture and flat area.Finally,the paper proposes a non-reference stereoscopic image quality assessment algorithm based on sparse representation of content for looking for the high-quality stereoscopic image quickly.The steps are as follows:firstly,using the left and right images to construct a single-eye diagram and an absolute disparity map,which reflect the competitive characteristics of binocular;then,sparsely representing different regions of different images by combining over-complete dictionary with image area division;finally,building a regression model by fuzzy depth confidence network to predict stereoscopic image quality.Experiments show that:on the one hand,the stereo image quality assessment algorithm presented in this paper is better than other methods that use the idea of sparse representation;on the other hand,the algorithm the paper proposes is more accurate than some full-reference methods.