Mutual information based registration of SAR images

Interpolation artifacts caused by the discrete nature of digital images and interpolation kernels have been previously reported in the literature for mutual information based image registration. Unfortunately they become pronounced in the case of SAR (Synthetic Aperture Radar) image registration, exacerbated by speckle noise. We analyzed the two widely used interpolation algorithms in image registration, namely bilinear interpolation and partial volume interpolation. According to simulation results, we found that the former algorithm produces an extremely nonsmooth metric function, whereas the latter method introduces spurious global optimum at some position of misalignment, instead of the position of perfect registration, under certain circumstances. To make the mutual information criterion useful for the application of SAR image registration, we proposed two pre-processing steps including speckle reduction and an appropriate selection of the number of bins for histogram construction. Simulation results indicate that, by this effort, interpolation artifacts in the mutual information function can be significantly suppressed, and as a consequence, accurate registration is allowed. In conjunction with the two proposed pre-processing steps, we implemented a multiscale elastic registration algorithm. An example of registering a pair of JERS-1 (L-band, HH-polarization) and RADARSAT (C-band, HH-polarization) images verified the potential of this algorithm.