When the wavelet decomposition is used to denoise corrupted natural picture, the precision of variance estimation of wavelet coefficient has great influence on the denoising effect. Because the wavelet coefficients in different scales have quite different features, it is necessary to choose different window sizes in different scales to estimate the variance of coefficient in the window center. The outcome of variance estimation by maximal likelihood criterion is firstly analyzed. Then the variation and power of wavelet coefficient are also considered to determine the proper size of window in different scales. At last several simulations are performed to test the above analysis. The excellent simulated result shows the power of choosing proper window size in the specific scale.