ABSTRACT
The traditional regularization terms previously discussed (depending on the local image derivatives) are based on local image operators (like the image gradient), which denoise and preserve edges very well, but may induce loss of fine structures like texture during the restoration process. More recently, Buades et al. [72] introduced the nonlocal means filter or NL-means (as a type of neighborhood filter), producing excellent denoising results. Gilboa and Osher [154, 155] formulated the variational framework of NL-means by proposing nonlocal regularizing functionals using nonlocal operators such as the nonlocal gradient and nonlocal divergence. These have been applied very successfully to image denoising, inpainting, and restoration [154, 155, 261, 223, 183].
