Fully non-local super-resolution via spectral hashing

Abstract

Super-resolution is the task of creating an high resolution image from a low resolution input sequence. To overcome the difficulties of fine image registration, several methods have been proposed exploiting the non-local intuition, i.e. any datapoint can contribute to the final result if it is relevant. These algorithms however limit in practice the search region for relevant points in order to lower the corresponding computational cost. Furthermore, they define the non-local relations in the high resolution space, where the true images are unknown.In this work, we introduce the use of spectral hashing to efficiently compute fully non-local neighbors. We also restate the superresolution functional using fixed weights in the low resolution space, allowing us to use resolution schemes that avoid many artifacts.

Adaptive Regularization of the NL-Means: Application to Image and Video Denoising

“…Protter et al [11] and d'Angelo and Vandergheynst [12] use a non-local data fidelity term combined to TV minimization in order to obtain a highquality image from a low-resolution image sequence. Their formulation shows some similarities to our proposed model, but the philosophy behind it remains quite different.…”

“…R-NL offers a simple way to deal with this compromise, thanks to the sum of the non-normalized weights that acts as a measure of confidence in the denoising performed by the NL-means. To illustrate further the difference of philosophy between our R-NL method and the algorithm derived from the super-resolution context [12], Fig. 3-g) displays the difference between the solution of the NL-means and the solution of the algorithm derived from [12], while Fig.…”

“…To illustrate further the difference of philosophy between our R-NL method and the algorithm derived from the super-resolution context [12], Fig. 3-g) displays the difference between the solution of the NL-means and the solution of the algorithm derived from [12], while Fig. 3-h) displays the difference between the solution of the NL-means and the solution of our R-NL algorithm.…”

“…Protter et al [11] and d'Angelo and Vandergheynst [12] have used the same type of approach in the context of superresolution. Their technique could be adapted to a denoising context.…”

“…This sums up to computing the solution of the NL-means, then applying TV regularization. In this article, contrary to [11] and [12], we exploit weights that are not normalized. This allows us to use the weights to ponder the TV regularization, hence creating an adaptive model.…”