Fusion Algorithms Of Remote Sensing Images Based On Multi-resolution Analysis

The research objects are the multispectral and the high space-resolution panchromatic images attained by different sensors in the same satellite, and the purpose is finding appropriate algorithms which can be used in image fusion to make the fusion images preserve the spectral qualities as much as possible while improving the space resolution quality. In other words, there are two targets of fusing MS image and Pan image, one is preserving spectral qualities and the other one is improving space detail information. The whole research is according to the two main technologies, which are the IHS transform and image multi-resolution analysis, and here the image fusion arithmetics based on Mallat wavelet transform,àtrous transform andàtrous-Contourlet transform respectively are researched.The basic principle of fusion of remote sensing images based on IHS transform and multi-resolution analysis is as follows. First, MS images are transformed from the RGB space to IHS space through IHS transform, and then the intensity, hue and saturation are attained. Second, the intensity and the panchromatic image are transformed in multi-resolution analysis algorithm, acquiring the approximate information and high frequency detail information. With different fusion algorithms are used, the multi-resolution representation of the fusion intensity of the fusion image is acquired, and then the intensity is acquired through contrary multi-resolution analysis. The last, the fusion image is acquired in contrary IHS transform. There are two key probrems in the remote sensing images fusion, one is the multi-resolution decomposition of images, and the other is the selection of decision strategies.Wavelet, which is of good multi-resolution features in the time and frequency fields, just as the magnifier which has different multiples, can extract the details of the images layer by layer. Due to the features of un-redundancies,multidirectional detail information representation and fast algorithm, Mallat wavelet is paid great attention to in image fusion, based on which a lot of fusion methods are developed. After summarizing and comparing, a new gradient selection fusion algorithm based on Mallat wavelet transform is developed. The low frequency information of the intensity of MS image is selected as the low frequency of the fusion image to pretend the spectral quality as much as possible. To the high frequency, the gradient is used as activity measure, correlation coefficient as the match degree. The weighted fusion algorithm based on the window is developed. As proved in theory and experiments, the algorithm has good capability, and good fusion results can be acquired.But because of the sampling process of the Mallat wavelet transform, each frequency part has different size, only the wavelet substitute and weighted method can be used. To the fusion of MS image and the Pan image, the intensity change much, which affects the fusion results'spectral quality and space quality.àtrous transform is another wavelet multi-resolution analysis, different from the Mallat wavelet transform, it is un-sampling discrete wavelet transform, which can acquire low frequency and high frequency information of the same size as the original image. To increase the space resolution of the MS image, there is a way of extracting the detailed spatial information not present in an MS image from a Pan image, and later the detailed spatial information is injected into the MS image. In other words, the detail information which is acquired from two images can be added into the MS image directly, which needs the high frequency information layer has the same size with the original image. As of the reason,àtrous wavelet transform is an ideal multi-resolution analysis method for the fusion of the Pan image and the MS image. J. Nú?ez had developed the famous algorithm-WA.Here, the advantage and disadvantage of WA algorithm have been analyzed in theory, and an improved WA is developed. The basic principle is as follows: First, the intensity of the MS image and the Pan image are transformed inàtrous wavelet transform. Second, the weighted fusion algorithm based in window isused in the corresponding high frequency information of the two images. In the algorithm, the gradient is selected as the activity level measure, because the gradient can reflect the high frequency information sensitively, correlation coefficient as the match degree, in addition, to improve the space quality as much as possible, different principles are used in the different conditions of coefficients. The last, the fusion high frequency layer is added to the intensity of the MS image directly. It is proved that the wavelet addition algorithm based onàtrous wavelet transform is better than the wavelet substitute based on Mallat wavelet transform. In addition, better spectral quality and space resolution quality can be acquired in the improved WA algorithm developed here.The wavelet transform is a good multi-resolution analysis instrument, but it shows that the"curves"in the picture is captured in the"dots"finally, and it only can reflect the features of"crossing"the edges but not"along"the edges, so the detail information can not be described sparsely. As the development of the multi-resolution analysis, the Contourlet transform is developed by Minh N.Do and Martin Vetterli in 2002, in which the base function with the shape of rectangle is used in expressing the images, so the smooth contour can be described in a small quantity of coefficients. The Contourlet transform, which is also named pyramid directional filter bank, is realized as follows: First, the image is decomposed in Laplacian pyramid, and then to the high frequency layer, the directional filter bank is used to acquire a Contourlet high frequency coefficient from the same directional coefficients. So through the transform, the high frequency and the low frequency are of different sizes, just as the Mallat wavelet transforms. Because of that, both of the theory and the experiment have show that it can not acquire better results than WA algorithm, so to take the advantage of the Contourlet transform abundantly and the good features of theàtrous wavelet transform, some changes are made to the Contourlet transform. The change is that the pyramid decomposition is replaced by theàtrous wavelet transform, thenàtrous-Contourlet transform is proposed.In this transform, there are advantages of both theàtrous wavelet transformand the Contourlet transform. The transform is realized in the processes as follows: First, the image is transformed inàtrous wavelet transform and the high and low frequency layers of the same size can be acquired. Then the directional filter is used to the high frequency layer, and the high frequency information of the same resolution is acquired. For the image fusion, similar to the other kinds of multi-resolution analysis, the high frequency of different directions is fused based on different fusion algorithm, and the high frequency direction information of the fusion image is acquired. Then the high frequency layer of the same size as the original image can be acquired in contrary directional filtering, which contain more high frequency information. The last, the high frequency is added to the intensity of the MS image in WA algorithm. Here, to prove the advantage of theàtrous-Contourlet multi-resolution analysis, some fusion algorithms are used, such as larger selected based on the pixels, and then a new algorithm-NWGSA is developed. It is proved that fusion algorithm based onàtrous-Contourlet transform can acquire much better space quality with the similar spectral quality as the algorithm based onàtrous wavelet transform.As all of the above, the image fusion algorithms based on multi-resolution analysis such as Mallat wavelet transform,àtrous wavelet transform andàtrous-Contourlet transform proposed here to the fusion of the MS image and the Pan image are researched. Series of experiments have proved that some new algorithms have good performance. In addition, theàtrous-Contourlet proposed here is a good multi-resolution analysis method for the remote sensing fusion, which will be of great importance to the research and engineering application.