In this paper, a novel clutter removal method based on non-local means (NLM) filtering for ground-penetrating radar (GPR) is presented. NLM filter which can be considered as a generalization of bilateral filter diverges from other local averaging filters since it determines the pixel weights by investigating the self-similarities in the image. NLM filter is extended to a multiscale-multidirectional version called multiscale directional non-local means (MDNLM) filter. Then, it is used to decompose the GPR image into approximation and detail subbands to capture the intrinsic geometrical structures of GPR image that contain both target and clutter information. After directional decomposition, the clutter is eliminated by keeping the diagonal information as target component. Finally, the inverse transform of the remaining subbands provides the reconstructed clutter-free GPR image. Results for both simulated and real datasets are presented to validate the superiority of the proposed method over widely used clutter reduction methods, as well as the recently proposed neighborhood filter based one, in terms of receiving operator characteristic (ROC) curves, thus improving detection performance.