Unsupervised microscopic image denoising network from single noisy images using similar adjacent patch matching

Abstract

Recent denoising methods based on deep neural networks commonly require a paired set of a clean image and its noisy version for training. However, in the microscopic or biomedical images, collecting the pairs is difficult, possibly requiring a human-in-the-loop process at an expensive cost. To overcome this problem, Noise2Noise (N2N) and Noise2Void (N2V) have been proposed to train the denoising network by only using noisy images. Nevertheless, N2N is applicable in a limited condition because it requires a noisy image pair from the same scene, which is difficult for dynamically living cells. In the case of N2V, the model only handles pixel-wise independent noise, so it cannot denoise a wide-field microscopic image that has noise appearing in multiple pixels. In this thesis, we introduce an unsupervised distribution-free denoising network, subNoise2subNoise (sN2sN), which does not require any noisy image pairs or assume pixel-wise independent noise. Our approach is based on the observation that a small patch of the clean natural resemble the adjacent patches. Our model is trained to make patches of the output to resemble each other and denoises an entire image at once during the inference time. We compare our method to the existing unsupervised denoising models on wide-field microscopic images and show that sN2sN quickly erases pixel-wise dependent noise that could not be denoised in the previous methods.