The fixed point iteration is employed when we restore the image by total variation, and we propose to select the regularization parameter by variance of noise in the iteration. The initial noise statistical properties in observe image is known hypothetically. In order to correctly estimate the variance of noise in iteration, a pure synthesis noise as an image is synchronously iterated with the observation image in de-convolution, and we take variance of pure noise image as the estimation of the variance of noise in observation image to compute the regularization parameter by the variance. Against the anisotropy of the regularization, the novel regularization term that can ensure the synchronous changing of the statistical properties of two noises was propose in this article. The new regularization term is put into use only in iteration of pure noise image, and the similarity of statistical properties between actual image noise and pure noise can be maintained in iteration. Under the condition of knowing the variance of noise of image in iteration, the relationship between the variance of synthetic noise and the regularization parameter λ can be established. The adaptability of total variation based image restoration is dramatically improved.