Research On SAR Image Registration And Fusion Based On Multiscale Analysis

Synthetic aperture radar becomes an important detection device in the modern military reconnaissance with its features of all-weather and diversified polarization and high resolution. However the single SAR imaging information can not meet the requirements of the integrality and accuracy of the target. As the development of the multi-source image registration and fusion, the optical imagery with the high identification degree and the infrared imagery with the ability to reveal high temperature camouflage have been becoming the strong complementary information. This paper aims at the speckle suppression and the registration of SAR and optical image and the fusion of SAR and infrared image on the basis of multiscale analysis. The main tasks are as follows:Focusing on the over-smoothing problem of the global filtering in the speckle suppression, this paper proposes a method based on curvelet transform and non-local filtering. The method transforms SAR image into curvelet domain and extracts low subband coefficients with more energy by non-local filtering. It can not only retain edge information, but also have better speckle suppression on the uniform areas. The test data experiment proves the method have solved the problem of edge distortion seriously in speckle suppression and reflected image contrast preferably.Focusing on the low efficiency of parameter iterative and the stability effected by false corners, this paper proposes a method based on individual optimal selection constraint. The method transforms optical image into wavelet domain and extracts effective corners. Then it transforms the corners into SAR image via pixel migration, and optimizes the parameters via the similarity measure criteria of gradient sum of squares and adopts polynomial to interpolation and resample. The experiment proves the method have reduced the number of false corners and increased the efficiency of iterative.Focusing on enhancing the common information and reducing false targets and complementing the background information, this paper proposes a method based on the curvelet transform and adaptive weight fusion. It transforms images into curvelet domain and adopts neighborhood Gaussian statistical model to the low coefficient. After that it adopts the strategy of the absolute bigger value to the high coefficient. The experiment proves the method can give full play to curvelet transform in the interpretation of texture information by multiscale direction and retain more structural information.