Research On Multiband SAR Image Registration Based On Feature Fusion

Compared with traditional optical image registration, Synthetic ApertureRadar (Synthetic Aperture Radar, SAR) image is not affected by theweather，and it is capable of all-weather, day-or-night working. MultibandSAR image registration can make full use of the different satellite SARimage data, it is of great significance. This paper research on multibandSAR image registration, the main contents are as follows:First of all,for SAR images, speckle noise suppression algorithm isstudied.This paper introduces the typical statistical characteristics of filteralgorithm based on statistical area, including Lee algorithm, Frost, Kuanalgorithm, etc. The purposeofthesefilteringalgorithm is to reduce noise andkeep characteristics of edge and target.To accomplish this goal, so I introduceL0norm filtering algorithm, the algorithm’s goal is to complete the gradientminimization. L0norm filtering can effectively remove the noise andpreserve important edges.The paper based on L0filtering and the fact thatinput images’ grayscale characteristics should be similar for imageregistration, propose improved filtering algorithm.After Multi-mode SARimage filtered by improved L0filter, the quantity and percentage of matchedSIFT features obtain aimprovement, advancing the efficiency of the SIFTfeature extraction and matching.Second, in the matching stage,combining the phase correlation andASIFT features, I proposea new feature fusion registration process.Fouriermellin transform based on phase correlation and imageregistration parametersare calculated by cross-power spectrum inverse transformation. But thismethod depends on the size of the public area and has lower success rate ofmulti-mode SAR image.Introduced with affine invariance ASIFTcharacteristics, When Fourier mellin transform registration failed, the inputimages experience ASIFT feature extracting and matching after L0filtering,and use RANSAC algorithm to eliminate points which does not conform tothe transform characteristic of the model.The whole process will be fusion bythese two features, improving the robustness and accuracy of the registration.Finally, according to image local registrationproblem,put forward amethod to have smooth control points after image coarse registration.Captureimage blocks from image window around the control point in two images, then these blocksare applied to template matching.After the templatematching,we get more accurate control points.Thin plate spline functioncorresponding transformation’s parameters are calculated by using the controlpoints. According to theseparameters, transform the coarse registered imageand complete the local image registration. The image registrationprocess hasbeen finished.