By modeling orthogonal bandelets coefficients in each quad-tree subsquare as Generalized Gaussian Distribution model
a calculation formula for adaptive local subsquarewise threshold is derived under the Bayesian frame
and the best range of the parameter needed to calculate subsquarewise threshold is found out. On these basis
a subsquarewise threshold denoising algorithm for natural images is proposed in bandelets domain. Owing to making full use of local statistic information of the image
the visual effect and evaluation criteria of proposed algorithm for natural image denoising outperform that of BayesShrink and other threshold-based methods.