Automatic Registration of Multispectral Images on Graphics Processing Units

The increasing programmability and parallelism of commodity graphics processing units (GPUs) makes them strong candidates for addressing some of the computational challenges faced by the remote sensing community. We have addressed one of those problems: the implementation of a fully automatic system for registration of remote sensing images. Our previous research in this field has focused on feature based methods. However, the massive parallelism offered by modern GPUs has driven us to explore an alternative similarity approach based on Differential Evolution (DE). Essentially, this method seeks the affine transformation that best align the target images – the transformation that maximizes the mutual information (MI) between them –. In our experiments, DE provides accurate results but unfortunately, it is computationally expensive. This latter aspect compromises the feasibility of this method since most applications in this context require timely responses. Although DE is a highly parallel method that fits well in PC clusters and other convectional parallel machines, computing the MI in the GPU is a challenging task since it involves 2D histogram calculations. In this paper we revise the major issues involves in this mapping.