Fast Fourier Convolutions in Self-Supervised Neural Networks for Image Denoising

Abstract

Quality of digital images depends on a multitude of environmental and equipment factors. In many cases our options for optimizing imaging conditions are limited, and the acquired images turn out to be corrupted with noise. Recently, denoising convolutional neural networks (CNN) have started to outperform classical denoising algorithms. If approached naively, these networks require a lot of pairs of noisy and clean images from the particular domain. In some fields (e.g. in biomedical imaging) it is hard to collect such data in abundance. This limitation has accelerated a research for self-supervised networks what can learn denoising just from noisy images alone. However, such networks’ performance could be constrained by the the limited receptive field of regular convolution. To mitigate this problem, a new modification for CNNs was proposed: Fast Fourier Convolution (FFC). Here, a global receptive field is achieved by using Fourier Transform and convolving spectral representation. Global perception field can help CNNs to better capture dependencies in image regions which are far apart. Given the ability of FFC to enhance multiple state-of-the-art classification neural networks, we hypothesize that denoising neural networks could also gain from its use. In this work, we design multiple approaches for incorporating FFC into self-supervised neural networks for image denoising. We evaluate these approaches on three diverse benchmark datasets and compare them with both supervised and self-supervised methods. We empirically show that FFC-enhanced denoising network achieves the state-of-theart results on character dataset and shows comparable level of performance for both grayscale and color natural images.