Total variation (TV)-based image denoising has still remained an active research filed of image processing. In the proposed model how to describe local structures of image is very important to improve the restored quality.This paper proposes an adaptive direction TVpimage denoising model, where the gradient operator can efficiently depict local structures by coupling the rotation operator and the weighted operator. The adaptive angle θ(f) used in the rotation operator via the orientation field estimation mainly depends on the average phase angle of pixels with a suitable neighborhood window, so this choice is more reasonable to express the local structure information. Since the proposed model is nonsmooth and non-Lipschitz, we employ the ADMM to solve it by using the half-quadratic scheme to solve the related ℓ2− ℓ^p subproblem. We prove the convergence of the half-quadratic scheme under the framework of the alternating direction method with a gradually decreasing smooth parameter. Furthermore, we also give some discussions when using the ADMM to solve our proposed model. Some numerical comparisons with the classic TV-based models illustrate the good performance of our proposed model for the image denoising problem