Due to the complex sources and serious interferences of the infrared image noise, and the traditional wavelet threshold methods for estimation have large deviations. An infrared image denoising method based on threshold estimation of bidimensional empirical mode decomposition (BEMD) sub-band is proposed. The noisy image is decomposed by BEMD into bidimensional intrinsic mode function (BIMF) subbands, Gaussian mixture model is used to calculate noise variance of each sub-band. As the noise estimation only considers the noise components, effects of the feature components are reduced, and then the more accurate threshold is obtained. The adaptive threshold is set to filter the noise. Experimental results show that the proposed method avoids the disadvantage of the hard threshold function and the soft threshold function, the image is relatively clear and visual effects are improved. Compared with traditional denoising methods, its mean square error (MSE) is less than the other methods, and the peak signal to noise ratio (PSNR) increases by 0.5 dB～3 dB. The new method has a better denoising effect, and the more noise variance the more advantages can be obtained.