To distinguish image features from high level noises, we propose to exploit the spatial correlations between wavelet coefficients by replacing the thresholding process with a diffusion process. The proposed smoothing method is more robust to noise when applied to highly corrupted document images.