Abstract

Super-Resolution for Aliased Images

In this talk we discuss different techniques for training the super resolution algorithm with the focus on producing high quality results on the highly aliased and noisy data. Though such artifacts are not often present in the real images, they frequently appear in gaming and animation images and videos....


Session No: SIL8154
Speaker: Ahmad Kiswani
Type:

Developer Tools and Profesional Visualization

Date: Thursday - October 18, 2018 03:00 PM - 03:45 PM
Location: Hall D
Topics: Deep Learning and AI, Professional Visualization
Industry: General

In this talk we discuss different techniques for training the super resolution algorithm with the focus on producing high quality results on the highly aliased and noisy data. Though such artifacts are not often present in the real images, they frequently appear in gaming and animation images and videos. To solve this problem, we have introduced a special down sampling procedure that allows the network to simultaneously do anti-aliassing, denoising and super-resolution. Our experiments show that even though our model was trained solely on natural images, it is able to achieve significant improvement for aliased and noisy animation images.