Stereo View Synthesis With Monocular Image Based On Deep Learning

With the rapid development of computer hardware,the calculation power will no longer limit the evolution of Deep Learning.Therefore,a large scale of computer vision research based on Deep Learning arises.However,in the field of stereo view,the research with Deep Learning is not comprehensive.And the result of traditional algorithm varies a lot due to the quality of data,the direction of optimization etc.Besides,from the perspective of application of these algorithms,lots of them require the depth of a scene,which is hard to acquire for individuals,since the depth hardware is not a widespread device.To get around the depth device,other algorithms demand a sequence of images with special conditions,which is also difficult to fulfill in natural environment.Here we propose a new algorithm to generate stereo view with only a single monocular image.First a new Neural Network is constructed based on ResNet to estimate the depth of a scene,then a sequence of images from multiple angles of view is rendered with the calculation of disparity.Finally a Generative Adversarial Network is used to complete the empty pixels of new view images.Compared with traditional algorithms,this framework works better among many of scenes,and a single image is easy to capture.Meanwhile,the estimation of depth is also more accurate than the state of art works.