Burst Imaging with Learned Continuous Kernels

Abstract

Burst imaging is a technique that consists of taking multiple images in quick succession and merging them into one output image. By aligning and combining data from multiple frames, we can increase resolution, attenuate noise, reduce motion blur and expand the dynamic range to obtain a higher quality image. In this thesis, we propose a method that learns continuous kernels to process and merge burst frames. We show that the learned kernels adapt to local image information and take advantage of sub-pixel sample location information to demosaic, denoise and merge the burst into a high quality output.