It is very important for inverse problem to choose regularization parameter properly. With too little regularization parameter λ , reconstructions are too smooth. Conversely, with too much regularization parameter, the reconstructions have highly oscillatory artifacts owing to noise amplification. Currently, there are several regularization parameter selection methods for Tikhonov regularization problems such as the discrepancy principle and the generalized cross-validation method, but it mainly is for linear problem. Our main contributions are as follows. Firstly, we prove the principle of determining regularized parameters for nonlinear total variation problems. Secondly, a new adaptive parameter selection method for the nonlinear total variation regularization is proposed. In the proposed method, we use a fast total variation method for image restoration, then employ the restored image to estimate the regularization parameter. Experimental results show the proposed algorithm is very efficient and the quality of recovered images by our proposed method is competitive with other image restoration methods.