Hybrid Method for Biomedical Image Denoising
Andrey Serdjevich Krylov, Valeriy E. Karnaukhov, Nikolay Mamaev, Alexander Vladimirovich Khvostikov · 2019
Convolutional neural networks (CNN) show very good performance and achieve impressive results in 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. Thus new more reliable hybrid denoising methods were suggested. They include combinations of CNN and "classical" algorithms like Non-Local Means, BM3D, Bilateral, Anisotropic diffusion, Total Variation (TV), etc. However hybrid methods need non-reference automatic parameters estimation for classical algorithms. In this paper we present a hybrid DnCNN+BM3D method with automatic choice of the strength parameter for BM3D method. To control biomedical image structures by multiscale ridge based approach we analyze presence of regular structures in the ridge areas at the difference between noisy and filtered images. Test results for retinal image dataset DRIVE, natural images BSD and for real CT image show practical applicability of the method.