Image Registration Method Based On Non-separable Wavelet Decomposition

In the problem such as multi-sensor image fusion, we must to finish the image registration in front of the fusion image. Since the imperfect calibration image platform inconsistent physical location of the same scene from different sensors or other reasons the image generated by the image sensor, what results in an image that may exist between the relative translation, rotation, scaling. Accurately, image registration is to matching and overlay two or more images, which produced in different time, different sensors or different conditions. The main process of image registration technique includes four steps. First of all, it is to get the feature points by the feature extraction from two images. Secondly, it’s to find the matching feature point pairs by the similarity measure. Thirdly, it’s to get the image space coordinate transformation parameters by the matching feature points. Finally, it’s to finish the image registration by the image space coordinate transformation parameters.At present, wavelet transform in image registration has been used very widely. Among them, the more effective method is that, to use the tensor product of image wavelet decomposition approximation components and image registration method, which is on the basis of planar contour matching by using the shape-specific points. The main process of this method is to convert original image registration problem to a tensor product wavelet decomposition approximation components after registration problems. And it’s very effectively to save three-quarters of calculate the registration parameters computation. At the same time, it can be used for real-time image registration and has good effect of registration.Tensor product wavelet only emphasize on the edge of the horizontal and vertical direction. Non-separable wavelet is a real two-dimensional wavelet, it is isotropic, can extract the edge of the image in all directions, and can obtain relatively complete outline of the image. When this kind of characteristics is applied to the image registration, it can acquire the accurate positions of the invariant points of the affine transform. Based on this characteristic, this paper proposes an image registration method through calculate the registration parameters of the high-frequency sub-images of the non-separable wavelet decomposition of the original image. This method can transform the registration of the original image into the registration of its high-frequency sub-images. First of all, the correctness of the proposed registration method is proved according to the fast algorithm theory of wavelet decomposition. Secondly, a four channel non-separable wavelet filter bank is constructed and the registration steps of the method are given. The experimental results show that the method has better registration results. The amount of computation for computing the registration parameter is less than a quarter of the registration parameters computation of the original image. When compared with registration method based on the tensor product wavelet decomposition, the method has higher registration precision.