A Regularization Method For Image Restoration Without Knowing Noise Level

Image restoration is one of the basic problems in digital image processing. In recent years, the technology is widely applied in mang fields, such as radio astronomy, satelite remote sensing, medical imaging and industy vision. The usual methods include regularization method, iterative method, stochastic method and so on.In this thesis, an improved combining regularization framework of two parameters type, proposed by R.Youmaran and A.Adler, is presented and discussed for stable image restoration without knowing noise level of the input data. Some details of the improvements include the adoption of well-known L-curve criterion with a speed-up technique for a rough initial value of regularization parameter, the translation of Euler equation requiring very large amout of storage to an equation requiring much less memory, the related analysis and proof, the design of the new algorithm and so on.The theoretic analysis and numerical experiments results have shown that our new algorithm has fast convergence speed and good restoration effect, and values of ISNR are improved upon 0.3 db generally. We achieve all the numerical experments on persenal computer in Matlab language and wish to give a hand on this works.