Noise2Split — Single Image Denoising Via Single Channeled Patch-Based Learning

Abstract

The prominence and popularity of Image Denoising in medical image processing has been obvious since its early conception. Medical Image Denoising is primarily a significant pre-processing method for further image processing steps in various fields. Its ability to speed up the diagnosis by enhancing the sensory quality of noisy images is proven to be working in most of the cases. The efficiency of the deep neural networks for Medical Image Denoising has been well proven traditionally. Both noisy and clean images are equal requirements in most of these training methods. However, it is not always possible to procure clean images for various applications such as Dynamic Imaging, Computed Tomography, Magnetic Resonance Imaging, and Camera Photography due to the inevitable presence of naturally occurring noisy signals which are intrinsic to the images. There have been self-supervised single Image Denoising methods proposed recently. Being inspired by these methods, taking this a step further, we propose a novel and better denoising method for single images by training the learning model on each of the channels of the input data, which is termed as “Noise2Split”. It ultimately proves to reduce the noise granularly in each channel, pixel by pixel, by using Single Channeled Patch-Based (SCPB) learning, which is found to be resulting in a better performance. Further, to obtain optimum results, the method leverages BRISQUE image quality assessment. The model is demonstrated on X-ray, CT, PET, Microscopy, and real-world noisy images.