No-reference Stereoscopic Image Quality Assessment Based On Deep Learning

As one of the carriers of information,image has become the most common way of information dissemination in people's daily life,work and entertainment.However,in the process of image acquisition,compression,transmission and display,various types of distortion will inevitably occur,which will result in the degradation of image quality and affect the subjective experience of the human eye when viewing the image.Accurate and efficient evaluation of image quality has become an important research topic in the field of image processing.In recent years,with the rapid development of stereoscopic image display technology,people enjoy the stereoscopic image far beyond the2 D image viewing experience,meanwhile,the stereoscopic image quality requirements are more stringent.Compared with the 2D image quality evaluation,the stereoscopic image quality evaluation is just in its infancy.Traditional methods of quality evaluation of stereo images depend on artificially extracted distorted image features.However,due to people's insufficient understanding of the principle of stereo image perception,it is impossible to extract accurate features to construct a stereo image quality evaluation model,resulting in the current no-reference stereoscopic image quality assessment methods can't satisfy the actual demands.Nowadays,the more and more popular deep learning technology can automatically extract useful features from the original image by training,avoiding the shortcoming of inaccurate hand-crafted features.Based on convolution neural network,this paper presents a new method for evaluating the quality of non-reference stereoscopic images based on global and local features of images.In this paper,based on convolutional neural network,the proposed no-reference stereoscopic image quality assessment method considers the quality of images from the perspective of global information perception and local information perception,evaluating the effects of distortion on the overall semantic information of images and on local detail information.After that,the proportion of salient regions relative to the full image is utilized to weight the global and local scores,generating the overall imagequality scores.The input information of the model are the left and right views which are normalized sufficiently,and there are no other features extracted manually,avoiding additional noise effects.In this paper,the accuracy and generalization of the proposed algorithm are verified on the LIVE 3D quality database and IVC database.The experimental results show that the quality prediction values of the proposed algorithm achieve high correlation with the subjective quality evaluation values of human eyes.The proposed algorithm outperforms most current non-reference stereo image quality evaluation algorithms and even some full-reference image quality evaluation algorithms.Especially in the quality evaluation of asymmetric distorted stereo images,the performance of the proposed algorithm has been significantly improved.