Actual limits of non-local methods in using image self-similarity
Résumé
Recently, significant progress has been made in different image processing tasks by introducing non-local operators that claim to exploit self-similarity properties of natural images. We give several examples of such image processing tasks and we then focus on the image denoising problem and especially on the well known non-local means algorithm. We show that, at least in the case of this algorithm, the so-called non-local operator remains essentially a local one. We present some arguments to explain this behavior and we conclude that a true non-local operator able to exploit image self-similarity remains to be invented.