Research And Practice On CNN Based No-reference Stereoscopic Image Quality Assessment

Recently, with the rapid development of multimedia, image information has been widely used in our life. Images express abundant and intuition information, especially the stereoscopic images. They create a truly immersive experience, which could bring many new applications to the world. However, stereoscopic images may be damaged in the process of image gathering, compressing, converting, decoding, and display. The way to assess the images quality effectively may influence a lot in processing. Traditionally, subjective image quality assessment methods have been applied, which were always time-consuming and exhausting. On the other hand, limited by the structure of stereoscopic images, the objective methods based on computer technology had shown less generality and accuracy. Because of these imperfections, we desperately need a new method which could assess the stereoscopic image quality by computer automatically and effectively.Along with the progress of neural network, deep learning, especially the convolutional neural network (CNN), has made the breakthrough in object identification, image recognition and some other fields. The CNN model could exploit the deep information in massive amounts of data, extract characteristic information, and solve the target problem by self-learning. Considering its powerful feature, we applied CNN algorithm to build a no-reference model. We employed 2D images and stereoscopic images to assess the stereoscopic image quality. The detailed work and innovation points of this paper are as follows:Firstly, taking its advantage of the high-efficiency in processing image information, we built a CNN model to assess the stereoscopic image quality. Through the abilities of learning the image structure and self-training, combine structure learning and objective training as a whole, this model could construct automatically without source images. As a result, this model is applicable to a variety of situations.Secondly, this paper analyzed the structure characteristics of the stereoscopic images and built three different kinds of CNN’s to exploit deep information of these images from various aspects. Making full use of multi-channel effects and stereoscopic perception of human visual system, we analysis the feature information from left and right view image and difference image, and then establish linear regression to analysis these information, finally match the subjective quality.Finally, based on the research on different CNNs, the problem of insufficient of stereoscopic image data is solved. By processing the differential correction data and visual information in stereoscopic images while taking advantage of similarities between them, we made use of transfer learning to apply 2D image quality assessment to stereoscopic images. In this way we developed the ability of network learning and generalizing based on expanding our database.Our model has been tested on LIVE 2D, LIVE 3D, and IVC database. It has shown higher consistency than current models no matter under same datasets or different ones. As a result, our model has improved the shortcomings of wasting of resource and low efficiency. This method has reference value and practical significance. It will promote the development of evaluation method of stereoscopic images quality...