Researches On Remote Sensing Image Denoising And Fusion Based On Support Vector Machine

During acquisition and transmission, remote sensing images are often corrupted by diversified noises, which could have a negative impact on the following process such as image segmentation and target recognition. Image denoising is to eliminate noise and remain the original image information as much as possible. To overcome the disadvantage of information insufficiency in the remote sensing images from single sensor, remote sensing image fusion is to conduct space registration on the same area image data originating from different sensors, adopt certain algorithm to combine information advantage or complementarity in each image data organically, and achieve an image with more abundant information.Support Vector Machine(SVM) is a new method of machine learning. Based on the statistical learning theory, it can solve "small" sample problem well, and have some advantages, such as superior performance in signal-to-noise ratio, high machine-learning efficiency, and widely applicable to other applications.In this thesis, researches on remote sensing image denoising and fusion based on support vector machine are executed and the main work are as follows:1. The actuality of remote sensing image denoising and fusion were analysed. The principle of support vector machine was introduced, and implementation of SVM was particularly discussed.2. Remote sensing image denoising based on support vector machine was studied, and support vector value filter by support vector regression model and nonsubsampled directional filter bank were applied to SAR image denoising. Based on support vector machine's advantages of strongly catching singularity ability, a new despeckling algorithm for SAR image was proposed based on support vector value contourlet transform. By combining support vector machine with nonsubsampled directional filter bank, support vector value contourlet transform was constructed, which exploited the directional character of support vector value. Then the transform was used to decompose SAR image at multi-scale,multi-direction and local adaptive Bayes shrink factor was linked up to reduce SAR speckle noise. Experiments using plentiful real SAR images indicate that the proposed algorithm performs better not only on background smoothing but also on preservation of texture than wavelet algorithm. Both numerical index and visual effect are improved apparently.3. Remote sensing image fusion based on support vector machine was studied, support vector value transform and pulse coupled neural networks(PCNN) were applied to remote sensing image fusion. A novel algorithm of remote sensing image fusion was proposed based on support vector value transform and pulse coupled neural networks. By using support vector value transform, the input images were decomposed into a number of sub-images with various scales, and were extracted support vector value. Then, at different levels, spatial frequency in support vector value transform domain was input to motive PCNN, the fused coefficients could be generated. Compared with the fusion algorithm based on nonsubsampled contourlet transform(NSCT), the simulation results indicate that the proposed algorithm can effectively improve information and visual quality, enhance spatial detail of the fused image.