Hybrid Method for Biomedical Image Poisson Denoising
Valeriy E. Karnaukhov, Andrey Serdjevich Krylov, Yong Shan Ding, Mylène C. Q. Farias · 2020
Convolutional neural networks (CNN) have recently become the main tool for the problem of biomedical image denoising. Nevertheless, CNN-based methods strongly depend on the used training set, and even small differences in the input data can cause output disturbance. One of the possible ways to tackle this problem is to use hybrid denoising methods that include combinations of CNN and "classical" image denoising algorithms. In this paper we present a hybrid iterative DeepRed algorithm for Poisson and mixed Poisson-Gaussian denoising with an automatic choice of the number of iterations. The choice is based on the control of biomedical image structures. Here we analyze the presence of regular structures in the ridge areas at the difference between noisy and filtered images by a multiscale ridge based approach. It was also discovered that the application of the Anscombe transform in the hybrid method improves denoising results. Test results for retinal image dataset DRIVE and Set12 natural images show practical applicability of the method.