A kind of infrared image threshold denoising method is given. It combines wavelet transform with generalized cross validation. An asymptotically optimal threshold can be determined, without knowing the variance of noise, only using the known input data. After making discrete orthogonal wavelet transform to an infrared image, denoising is done in the high frequency subbands of each decomposition level respectively, so that the maximum signal-noise-ratio can be obtained in the high frequency subbands respectively. According to the experimental result ,the given algorithm can reduce the noise of infrared image effectively ,while it also keeps the detail information of infrared image well. As to performance and visual quality, the algorithm is better than the wavelet thresholding given by Donoho, the modified wavelet thresholding given by Johnstone and the traditional median value filtering method.