Adam Schmidt MS Project Defense

Wednesday, June 12, 2019 - 1:00pm to 2:30pm
HFH 1132
Fast Rendered Image Denoising Using Adaptive Blockwise Computation
Adam Schmidt
Pradeep Sen (Chair), Lingqi Yan


This project looks at solutions for removing noise from rendered images in real-time. Rendering is the process of creating an image from a model of a scene or environment. Creating detailed renders can be an extremely costly process, so a recent push in graphics has been to create noisy renders and then estimate the detailed output. Given a rendered image, how can we quickly denoise this result? Fast denoising would be useful when a game company wants to denoise the scenes in their game in real-time, or when an animation studio needs quick results to see how a scene will look in their final animation. This project focuses on multiple concepts that can help machine learning models approach denoising in real-time. The main novel concept the project introduces is an optimization to allocate computation based on region detail. We use a chain of multiple Convolutional Neural Networks (CNNs) to estimate a detailed image at different levels of refinement. After each refinement a binary-mask estimator is introduced to decide which regions have reached sufficient levels of quality. The model then continues on more complicated image regions. In my talk I will detail the model's results and give a brief survey on the current state of the art.

Everyone welcome!