To speed up computer graphics methods, we often reduce the amount of data. At ray casting, for example, this can be achieved by sending fewer rays to certain parts of the image. This results in scattered data from which we want to reconstruct the final image. This is done by solving a partial differential equation using iterative methods that solve a system of linear equations. The basic iterative methods have slow convergence, so we present a multigrid method that works on grids of different resolutions and thus achieving better convergence. Since very little data may be available at reconstruction, for example 5%, the final image does not contain the details of the original one. Therefore we develop a convolutional neural network with an autoencoder architecture, which allows us to partially recover the details of the original image.
|
