How to choose adaptively parameters of image denoising methods?

Abstract

The problem of adaptive choice of strength parameters for wide class of mathematical image ridge and edge preserving denoising algorithms is considered. It arises now in hybrid denoising algorithms containing a combination of convolutional neural networks (CNNs) and these classical methods. The problem is considered for the case of additive white Gaussian noise. We find the denoising method parameters to maximally suppress the image noise while retaining important image structures. Multiscale ridge based approach is used to analyze presence of regular structures in the ridge areas at the difference between noisy and filtered images. Hybrid methods using Deeply-Recursive Convolutional Network and Non-Local Recurrent Network are developed. CNNs are used in combination with total variation based method with adaptive parameter choice.