Denoising in a Jiffy: A fast and GPU friendly algorithm for image denoising

摘要

Non-local means is a state-of-the-art image denoising algorithm which uses weighted contribution of similar patches to denoise images. But its asymptotic complexity upper bounds the degree to which the algorithm can be accelerated. In this paper, we present an approximate version of the same which uses Locality Sensitive Hashing to reduce the complexity. Finally, we show that the new proposed approximate algorithm is magnitudes faster that the original algorithm while maintaining 90% of the denoising performance. A GPU implementation of the algorithm which obtains over 100 times speedup is also presented.