Non-local means versus non-local total variation
Résumé
The well-known non-local means denoising algorithm may be inserted in the total variation optimization framework, resulting in a non-local total variation model initially introduced in the context of image deblurring. In this talk we compare, in the context of image denoising, the original, spatial-domain, non-local total variation scheme to the reference non-local means and to the plain total variation algorithms. In order to better preserve strongly oscillating structures like microtextures, we propose to add a total variation term in the non-local Fourier domain so that to obtain a spatial-frequency denoising scheme. Following a simple yet effective technique, this scheme is applied on overlapping patches and the final estimation is computed by aggregation. In the end we obtain a new algorithm that compares favorably with most denoising methods, enabling our model to start playing in the "big league".