SAR And Optical Image Registration Based On Gradient Mutual Information

Image registration is a fundamental task in image processing used to match two or more pictures. It is widely used in medical image analysis, remote sensing, computer vision and precision guidance localization and so on. Multi-sensor images refer to the images produced by the sensors with different imaging mechanism. Compared to the single-sensor images, the image information provided by multi-sensor images has reliability,complementary and redundancy. Thus, people can get a more reliable, more comprehensive and accurate image description by the image fusion between multi-sensor images. Image registration is the basis of image fusion, the level of registration accuracy directly affects the quality of user's application.Multi-sensor images are in higher gray contrast and smaller correlation, therefore this paper proposes a new kind of measurement into SAR and optical image registration, which combines the gradient space information with mutual information together. In the process of registration, we not only consider the correlation of gray information, but also the correlation of spatial information, which leads to a good registration result. Simulation results show that, compared to the traditional mutual information registration method,when less multiplicative noise appears in synthetic aperture radar images, the gradient mutual information registration method has smoother curve with a sharp peak value, which is easy to find the best registration parameters. Therefore, the method can reduce the gray contrast influence, which improves the precision and accuracy of registration for multi-sensor images.However, the gradient mutual information registration is susceptible to noise interference.When much multiplicative noise appears in synthetic aperture radar images, registration function is easy to fall into local extremum, which results in wrong registration parameters.Based on the stationary wavelet transform, this paper applies a bayesian wavelet shrinkage method for SAR image despeckling. First, the statistical model of wavelet coefficients is analyzed, which derives a fuzzy shrinkage factor based on the minimum mean square error(MMSE) criteria with Bayesian estimation. Then, the ideas of regions divided and fuzzy shrinkage are adopted according to the intrascale dependencies of the wavelet coefficients.The noise-free wavelet coefficients are estimated finely. Through the experiment, it can be concluded that the method can preserve edges during despeckling, which improves the accuracy and robustness of SAR and optical image registration.When the image size increases, given the large amount of calculation of the gradient mutual information algorithm, in order to improve the efficiency of the image registration,this paper applies registration strategies from coarse to fine toward images with the method of Gaussian Pyramid decomposition. The images are first decomposed by Gaussian kernel function, which produces two levels of decomposition. Then our search strategy works from deepest level of decomposition to the top level decomposition, going from coarse to fine spatial resolution. At each level the search focuses in on an interval around the best transformation found at the previous level and is refined at the next level up. We use three groups of SAR and optical images to verify the effectiveness of the method above, the results show that the Gaussian Pyramid decomposition method yields significant improvement in terms of speed over gradient mutual information method, which results in a two to three times the reduction in registration time.