Denoising Image Gradients Using Deep Learning Approach for Cinematic Volume Rendering.

Abstract

The Monte Carlo integration methods for solving light transport rendering simulations have shown auspicious results in photorealistic rendering. Specially this rendering simulation is performed with Monte Carlo path tracer. The problem with this algorithm is the presence of noise in the rendered images for low samples. On the other hand, neural networks are widely used for image and video denoising which provides favorable denoising results. In this study, we present denoising method for cinematic volume rendering scenes in order to achieve photorealistic visualization. We extend Monte Carlo path tracer to gradient domain. We estimated image gradient using a machine learning approach and demonstrated a network that computes the horizontal and vertical finite-differences of the image gradient. Our network aim is to generate clean gradients from noisy gradients. Finally, by using a screened Poisson reconstruction algorithm we produced a noise free rendered image