Multi-sensor image registration by combining local self-similarity matching and mutual information

Abstract

Automatic multi-sensor image registration is a challenging task in remote sensing. Conventional image registration algorithms may not be applicable when common underlying visual features are not distinct. In this paper, we propose a novel image registration approach that integrates local self-similarity (LSS) and mutual information (MI) for multi-sensor images with rigid/nonrigid radiometric and geometric distortions. LSS is a well-performing descriptor that captures common, local internal layout features for multi-sensor images, whereas MI focuses on global intensity relationships. First, potential control points are identified by using the Harris algorithm and screened based on the self-similarity of their local surrounding internal layouts. Second, a Bayesian probabilistic model for matching the ensemble of the LSS features is introduced. Third, a particle swarm optimization (PSO) algorithm is adopted to optimize the point and region correspondences for maximum self-similarity and MI and, ultimately, a robust mapping function. The proposed approach is compared with several conventional image registration algorithms that are based on the sum of squared differences (SSD), scale-invariant feature transforms (SIFT), and speeded-up robust features (SURF) through the experimental registration of pairs of Landsat TM, SPOT, and RADARSAT SAR images. The results demonstrate that the proposed approach is efficient and accurate.