After analyzing an exponentially decaying inter-scale model of image wavelet coefficients, a subband-adaptive thresholding algorithm is proposed based on Bayesian rule. The new approach outperforms other methods because it captures the statistical inter-scale property of wavelet coefficients,and it is more adaptive to the data of each subband. Experiments show that higher peak-signal-to-noise ratio and better subjective visual effect can be obtained, which are better than those of other thresholding denoising algorithms.