Multisensor Image Registration Algorithm Combining Sift and Particle Swarm Optimization for Application in Multispectral Imagery

Registration of multispectral images with other sensor image such as optical, SAR images, which is the process of estimating the misalignment between two images, is a crucial preprocessing for many applications of multispectral images [1], such as fusion and change monitoring. Recently, some methods have been proposed for multisensory image registration in remote sensing [2]-[5], such as Pixel Migration method proposed in literature [5] is effective to find the correct solution. Unfortunately, well registration performance is severely limited to the situation when geometric deformation is not significant, while it is unable to obtain satisfactory results when the deformation is large. In addition, how to choose appropriate bands of multispectral images to be registered is an extremely important problem related to registration accuracy, while there is little relative work. In this paper, to solve the problems, a step-wise algorithm combining Scale Invariant Feature Transform (SIFT) and Particle Swarm Optimization (PSO) is proposed for multispectral image registration with other sensor images. Firstly, information entropy as a rule is used to select a suitable band image from the misaligned multispectral images for registration. Secondly, a SIFT modified by combining with local invariant moment is used to search optimally matched points for coarse image registration. Then, the parameters of the coarse registration model are used to initialize an optimization process controlled by PSO where local gradient is used as a rule for searching optimal parameters for final registration.