Anti-aliasing for real-time applications in 3D using deep convolutional neural network

Abstract

In this paper we present a convolutional neural network model for solving the long- standing aliasing problem in the real time 3D graphics industry. Aliasing refers to the problem of having hard jagged edges in the rendered scene. These jagged edges become a distraction and on a large enough amount, creates an unpleasant viewing experience. There are quite a few techniques out there to counter this problem, namely, FXAA, NFAA, DLAA. Our neural network architecture consists of two-dimensional convolutional layers and max pooling layers for reducing the spatial dimension. We then generate the nal output from transposed convolutional layer. Our model is trained on a specialized (trained on a per application basis) and generalized (trained on a variety of dataset to work on all possible conditions) version for anti-aliasing. Based on SSIM and PSNR scores we found out that a specialized version of our model works best, both in terms of visual score and image quality metrics.