Virtual Viewpoint Generation Based On The Compressive Light Field

Virtual view-point generation is one of the key technologies of the three-dimensional (3D) display, which renders the new scene image perspective with the existing viewpoints. The 3D scene information can be effectively recovered at different viewing angles, which allows the users to switch between different views. In this paper, the matching theory of virtual viewpoint generation is discussed, which can provide the theoretical support for the study of virtual viewpoint generation.Firstly, two kinds of virtual view generation systems based on four-dimensional light field are proposed. One is the free-view TV system based on ray space, and the other is the 4D compressed field system based on liquid crystal panel. The two systems are compared and analyzed. The virtual view generation system based on ray space involves matching, interpolation and compression, and it is easy to produce large error. The system based on liquid crystal panel provides a simple system platform, which can restore the original light field information and has some advantages.Secondly, a method of virtual viewpoint synthesis is proposed for 4D compressive light field formed by liquid crystal panel. Actual light field information rather than the geometric information is used. Moreover,for better making the data actual meaning, we mainly use the nonnegative tensor factorization (NTF). A tensor representation is introduced for multilayer display. The light field emitted by an N-layer, M-frame display is represented by a sparse set of non-zero elements which are restricted to a plane within an Nth-order, rank-M tensor. The tensor representation allows for optimal decomposition of a light field into time-multiplexed,light-attenuating layers using NTF. Finally, the compressive light field of multilayer display information synthesis is used to obtain the virtual view-point by multiple multiplications. Experimental results show that for the approach, not only the original light field is restored with high image quality, whose PSNR is 36.9 dB, but also the deficiency of traditional matching is made up, and any viewpoint can be obtained from N free viewpoints.