Image Denoising And Fusion Based On Directionlet Transform

Image denoising and fusion are two important components of image processing. During acquisition and transmission, images are often corrupted by diversified noises, which could have a negative impact on the following process such as image segmentation and compression. Denoising is to obtain clearer object and increase the recognition rate through filtering noises in the images. Fusion is to combine advantageous information from multiple images of the same scene to acquire more exact and comprehensive description of the image.Wavelet transform has found wide applications in image processing in recent years. Wavelet is the optimal bases for functions with point singularity. But in the case of high dimensional, wavelet analysis can not take advantages of the data geometrical features. Then it is not the optimal or the sparsest representation of the functions, so it can not make good use of direction information in images. To solve this problem, a series of new multiscale geometric analysis emerges to building the optimal representation of high dimensional functions, which have a wonderful prospect in the research of image denoising and fusion.This paper studies one of the multiscale geometric analysis tools, i.e. Directionlet, and its application on image denoising and fusion. The main innovative points are as follows:(1) We proposed an improved multiscale Directionlet transformation method. When the direction of the Directionlet bases matches that of the anisotropic object in images, Directionlet can represent images well, otherwise bases of Directionlet will degenerate into wavelets and have poor approximation power. In this paper, we find main directions of an image and construct the sample matrix adaptively, which can adaptively catch the anisotropic features in images.(2) An image denoising algorithm based on improved Directionlet is presented. For the high peaks and heavy tail character of Directionlet coefficients, we model each subband with a generalized Gaussian distribution, then denoise images combining the values of shape parameter and local variance estimate.(3) We also study the application of the improved Directionlet transform. As the improved Directionlet can catch image information from different directions, so it has better directionality and anisotropy. Image fusion with improved Directionlet transform and regional measures outperforms wavelet fusion method.This research is supported by National Nature Science Foundation (No.60672126, 60702062,60971128), National High Technology Research and Development Program (863) (No.2007AA12Z136), Major State Basic Research Development Program of China (973) (No.2006CB705707).