Research On Multi-view Content Generation Method Based On DIBR

People observe the scenes through their eyes in the real world.Under the influence of physiological factors and psychological factors,the images captured by the human eye give them a three-dimensional feeling.However,images are mainly transmitted in two-dimensional forms.Because of lacking depth information,two-dimensional images cannot give people a sense of three-dimensionality and immersion.In order to bring people more realistic experiences,three-dimensional display technology came into being.The three-dimensional display is divided into a visual-assisted three-dimensional display that require wearing the device and a naked eye three-dimensional display.Because wearing 3D glasses and other equipment is inconvenient,and long-term viewing will produce dizziness,the naked eye 3D has become a research hotspot.Both in scientific research and in commercial aspects,the naked eye 3D has a good future.It is necessary for the naked eye three-dimensional display to collect a multi-view image.However,multi-view transmission needs a lot of bandwidth and requires more storage space and computing power.With the virtual viewpoint generation method,the computing resources can be fully utilized and the demand for hardware resources is reduced.With the viewpoint generation algorithm,a small number of viewpoints can generate more viewpoints.At present,the virtual viewpoint generation method has become a hot research direction in the field of three-dimensional display.In this thesis,an algorithm for generating multi-view content using depth image rendering(DIBR)is designed and implemented.DIBR has the advantages of easy implementation and fast operation,and is widely used in virtual viewpoint generation technology.However,due to occlusion and other factors,the image generated by DIBR will have cracks and holes.The auto-encoder based on convolutional neural network is used to fix the image holes.Firstly,matching maps with different disparities are generated by the method of stereo matching,and then the matching maps generate the probability maps and color maps by two auto-encoders.Finally,according to the probability maps and color maps,the holes in the images are fixed.In order to enhance the effect of image restoration,multi-scale convolution kernel and residual network are added to the convolution neural network.The restored image has an improvement in both visual and PSNR compared to the image obtained with the classical algorithm Criminisi.Since the disparity map is needed in the virtual view generation,a disparity map generation algorithm based on convolutional neural network is designed and implemented.The PSNR value of the generated disparity map is higher than that of dispnet.