Research On No-Reference Stereoscopic Images Quality Assessment Based On Cyclopean Images

The quality of stereoscopic image not only affects people's perception of stereoscopic image,but also reflects the quality of stereoscopic image transmission,compression and other processing technologies.Therefore,it is necessary to design a reliable and efficient method to evaluate the quality of stereoscopic image.In order to simulate the human brain's perception mechanism of stereoscopic image,we design a fusion method of left and right viewpoints,and propose two no-reference objective methods for evaluating the quality of stereoscopic image based on the cyclopean image.Firstly,the disparity map is obtained by stereoscopic matching algorithm,which combines disparity map with cyclopean image to form enhanced image.Secondly,considering the different characteristics of images at different scales,the statistical features of natural scenes are extracted from cyclopean image and enhanced image by two kinds of Gaussian models at two scales.Then,a simple and efficient method is used to extract statistical features from disparity maps.Finally,the features is fused and sent into support vector regression,and the objective quality score of stereoscopic image is obtained by using subjective score as label.Secondly,according to the excellent ability of convolution neural network to extract features,a two-channel convolution neural network is designed to realize the quality evaluation of stereoscopic image without reference.Firstly,a convolutional neural network structure with dense connection network as the main body is established.Secondly,the cyclopean image is used as the input of one channel of convolution neural network and the input of another channel is disparity map,which plays the role of feature compensation.Then,considering the important role of disparity map in the formation of stereoscopic image,the weighted guidance of disparity map to cyclopean image is realized by modified the Squeeze and Excitation block.This weighting strategy enhances the proportion of useful information in cyclopean image and reduces the proportion of less useful information.Finally,at the end of the convolution neural network,the features of disparity map and weighted cyclopean image are fused to get the overall features,and the overall features and subjective evaluation scores are regressed to get the quality score of the stereoscopic image.Experiments are carried out on open databases.The experimental results show that the proposed methods can effectively deal with symmetrical and asymmetrical distortion stereoscopic images and maintain good correlation with subjective scores.