Abstract:During wavelet image processing, the variance of Gaussian white noise is usually estimated in the finest HH subband. A popular method, proposed by Donoho and Johnstone, is often found to provide too large an estimate. To tackle this problem, this paper presents a new method. The new method takes the rude estimate from Donoho’s method as the starting point, and then a subband more dominated by noise is produced with the signal filtered out by a filter derived from statistics theory and a newly-proposed coefRcient model, the doubly stochastic process. Thus a finer estimate is possible by using Donoho’s method on the filtered HH subband. Through employing EM algorithm, the new method can he straightly extended to the case of non-Gaussian noise. Experimental results show that the new method can improve the estimate quite much when compared to Donoho’s method.