Exploiting input space symmetries in video deinterlacing using radial basis function networks

Deinterlacing is the ability to estimate one field in a video image using the other field and as such is a form of interpolation. Such a task is required in many video applications, such as standards conversion. Unfortunately the performance of deinterlacers based on linear filters is unsatisfactory. In a previous paper, we have proposed radial basis function networks as a possible non-linear technique to address the problem. The focus of this paper is to exploit the symmetries that arise in the geometry of the input-output space originated from the sampling of the images. Since radial basis function networks create nonlinear mapping in this space, it is shown that these symmetries can be used to reduce the network's complexity.