Abstract: In order to remove the coherent speckle noise of SAR images, an improved Garrote threshold function is proposed, in which the exponential function is added to make it easier to approximate its asymptote, and the order of threshold parameters and independent variables in the function is increased to reduce its deviation. The improved Garrote threshold function is used in the wavelet threshold SAR image denoising method. Firstly, as the coherent speckle of SAR images is multiplicative noise, an normal method is used to convert the SAR image to additive noise models, so as to facilitate the wavelet filtering processing. Secondly, the pre-processed SAR image is decomposed to various frequency sub-band images by wavelet transform. The high frequency sub-band coefficients at horizontal, vertical and angular directions are denoised by improved Garrote threshold function and the low frequency coefficients are reserved. Thirdly, the denoised wavelet coefficients are used to reconstruct the image. Finally, the reconstruct image is sent to exponent arithmetic network, and the output of the network is the denoised SAR image. Equivalent number of looks and edge preservation index are used to analyse the experimental results, and the analysis shows that the noise of the image is removed effectively and the detail of the image is kept well.