To deal with the shortage of inaccuracy and low level robustness of block similarity measure for synthetic aperture radar （SAR） image despeckling using the traditional nonlocal means method, a new SAR image despeckling method based on adaptive nonlocal means is proposed. Firstly, the smoothness which is defined to measure image texture complexity is used to design the matching functions, and on the basis of those functions the size of image block and search window can be adjusted adaptively in order to improve the accuracy of block similarity measure. As a result, a framework of the adaptive nonlocal algorithm is proposed. Secondly, the block similarity is measured by the Gabor filter for enhancing the robustness of block similarity measure, and an adaptive nonlocal means method based on Gabor filter is presented when combined with the proposed framework. The experiment results show that the proposed method can not only reduce speckle efficiently, but also preserve the texture, edges and targets well, which lies the foundations for the understanding and interpretation of SAR images.