Contourlet coefficients are modeled using a hidden Markov tree (HMT) model with Gaussian mixtures that can capture all interscale, interdirection, and interlocation dependencies, which is more valid than wavelet HMT model. Put forward an improved contourlet-domain hidden Markov tree model where the state of the contourlet coefficients depends not only on the state of its parent node but on the state of the twin of its parent as well. This strategy can catch richer interscale correlation of the contourlet coefficient, and then is more suitable for representing non-Gaussian statistics and persistence of the contourlet coefficient. The improved model is used in infrared image denoising and compared with the other denoising method, such as wavelet threshold and wavelet HMT, and the simulation results show the method is more advantage restoring edges of original image.