Parallel-friendly patch match based on jump flooding

Abstract

In this paper, we propose a parallel-friendly algorithm for k-nearest neighbor based patch match. Based on jump flooding algorithm, an efficient pattern of communication, our algorithm is fully parallelized at patch-level. To improve the performance, we propose and analyze its variants, and implement them with GPU. Compared with state-of-the-art approximate patch match algorithm, the GPU implementation of our algorithm achieves up to 100 times speedup over its CPU implementation, and 5 times faster than the GPU implementation of Barnes's algorithm, a most recently benchmark algorithm. © 2012 Springer-Verlag Berlin Heidelberg.