Projection Conversion And Compression Of 360-degree Video

With the rapid development of virtual reality technology,panoramic video,as one of its important applications,is widely used in various fields.Panoramic video improves people's video viewing experience,brings great innovation and convenience to life,and also brings more opportunities and challenges to much related technology progress.This paper focuses on two very important steps in panoramic video processing: projection and coding.Aiming at the existing problems in the projection and coding process,it puts forward effective improvement measures or optimizes the existing methods.The work is summarized as follows:(1)Since there is still room to improve the performance of projection conversion between standard cube and sphere,we propose a post-processing method to improve the quality of panoramic video decoded by the client.Our method can be summarized as the combination of a convolutional neural network(CNN)and traditional interpolation to improve the transformation performance of cube to Equirectangular(ERP).Starting from the projection characteristics,we analyze the projection law of cube and sphere and deduce a formula to calculate the optimal boundary threshold of sphere pixels on ERP.At the same time,we use the Marsaglia model to propose a new evaluation criterion,which is easier and more accurate to evaluate the geometric transformation process.(2)Due to the problems of deformation,content discontinuity and disconnection at the boundary of geometric projection surface of panoramic video,we propose a padding method based on spherical interpolation.This padding scheme provides the boundary pixels have high-quality adjacent pixels,so as to improve the quality of boundary pixels during encoding.The padding pixels come from lossless sphere directly,which makes the padding and the content at the junction continuous and the pixel value and deformation degree more accurate;because the sphere is closed,and the interpolation method through the sphere is more accurate than the vertical processing,so our method is more universal and reasonable.The experimental results show that our algorithm has better performance in the subjective quality of the padding area compared with the comparative experiment and the reference algorithm,and has better objective quality improvement compared with HEVC(high efficiency video coding).(3)The main problem of panoramic video coding is deformation,but there is notechnology to predict panoramic video deformation in the existing methods.For this reason,we propose a panoramic video compression scheme based on spherical frame rotation prediction for the first time,adding a more accurate reference frame to predict the degree of deformation,so as to improve the panoramic video compression performance.Compared with the existing neural network frame prediction method for2 D video processing,the generation of rotation prediction frame is more convenient,accurate and fast;it can directly process panoramic video with original resolution;at the same time,our method can well fit the content of each frame;the scheme can well compatible with traditional video coding standards.The experimental results show that our rotation frame can predict the next frame deformation more accurately than the existing reference frame queue and can significantly improve the compression performance.