Noise Level Estimation In Curvelet Domain For Images

In this paper we study the variance estimation problem for noisyimage, which is corrupted by the zero mean additive white Gaussnoise. Assuming that image and noise are uncorrelated, the noisyimage g(x，y) can be represented asg(x, y)=f(x, y)+η(x, y)Where f(x，y) is the clean image, and η(x,y) is the additive white Gaussnoise with variance. This is one of the simplest and widely usedmodel in the digital image processing. The only parameter that need toestimate is the variance a~2.The noise signal corresponds to the high rfequency information ofthe image, but the main energy of high frequency information is not allprovided by the noise signal. For example, edges and textures alsocorrespond to the high frequency information of the image. Theseedges and textures will birng seirous effect to the noise estimation of theimage. How to distinguish these two kinds of high frequencycharacteirstics as far as possible in the image, overcome the influence ofother high frequency information such as edges and textures, andestimate the noise parameters accurately, is a very meaningfulproblem, which is also the center of this thesis.In this thesis, the estimation precision of the noise model isimproved by the following four aspects:1. Estimation of image noise in the domain of Curvelet transform;Curvelet transform is a multiscale pyramid with many direction andposition at each scale and needle shaped element at fine scale.Curvelet transform can sparsely characteirze the images which have line, curve or hyperplane singularities and the approximation eiffciency.The Curvelet coeiffcients of different directions reflect the differentdirection information of the original image. So，from the point ofdirection selecting, we can find the best direction in the curveletdomain for improving the accuracy of noise estimation.2.We construct the MVO (Maximum modulus value oirentationmatrix) and MVOCM (Maximum modulus value direction connectivitymatirx)，distinguishe the two kinds of high frequency structure: noisestructure and line structure of original image in the curvelet transformdomain effectively. Improve the accuracy of noise estimation from thepoint of set selecting;3.We prove the relation between the noise level and the size ofpoint sets when the estimation accuracy is given，as the theoreticalguidance for the noise estimation under different noise levels.4.We propose an adaptive neighborhood expansion technique toimprove the operability of point set selection for noise estimation.The experimental results show that our method works well forestimating noise at different levels.