An ideal guided image with good structure is the key to guided filtering (GF) denoising. In order to improve the denoising effect of the GF, an anisotropic total variation (ATV) based on the GF is proposed. First, the noisy image is smoothed by the ATV model to obtain a good structure information image. Then the image is served as the guided image in the GF algorithm. And an iterative processing is used for improving the algorithm robustness. Considering that the traditional iterative method for the total variation model is time consuming, the Split Bregman iterative method is introduced to speed up the whole process. The experimental results indicate that the proposed method not only has certain advantages in peak signal noise ratio, normalized mean square error and structural similarity, but also increases the computation speed by nearly 30 times compared with the related traditional iterative method. It can effectively remove the noises while preserving more structure and edge details.