Decentralized Multi-Agent Algorithm for Translational 2D Image Alignment

Abstract

We present a novel multi-agent algorithm applied to the problem of image alignment. Our method operates with multiple concurrent solutions held by agents who each attempt to reach the lowest error function score by trying to place a segment from a translated image to an unsegmented fixed image. Agents borrow and return segments of the translated image from a shared repository and iteratively suggest and evaluate their particular solutions. Finally, the global solution is determined by clustering of agents’ individual results. Experiments show that our approach provides results of high reliability and performance compared with traditional intensity based registration methods that rely on global optimization of a single error function given by translation of whole image.