Core Concepts
This paper presents the first fully unsupervised deep learning-based denoiser capable of handling imaging noise that is row-correlated as well as signal-dependent.
Abstract
The paper proposes an unsupervised deep learning-based denoising algorithm for noise that is correlated along rows or columns of pixels and is signal-dependent. The key aspects are:
The authors train a Variational Autoencoder (VAE) to model the distribution of noisy images. They design the autoregressive decoder of the VAE to have a 1-dimensional receptive field that can only model the row/column-correlated structure of the noise, but not the correlations in the underlying signal. This encourages the latent variables of the VAE to capture the clean signal content.
The authors then introduce a novel "signal decoder" network that is trained to map the latent variables produced by the VAE into an estimate of the clean underlying signal. This is done by using the original noisy images as training targets, similar to the Noise2Noise approach.
During inference, the authors sample latent variables from the VAE's encoder and use the signal decoder to produce denoised estimates of the clean signal. Averaging multiple such samples provides the final denoised output.
The method is evaluated on a range of real-world microscopy datasets affected by row-correlated and signal-dependent noise, as well as simulated noise datasets. It outperforms existing unsupervised denoising baselines, and even matches the performance of a supervised denoising method in some cases, without requiring any clean training data.
Stats
The paper does not provide any specific numerical data or statistics to support the key logics. The focus is on the methodological contribution and experimental evaluation.
Quotes
The paper does not contain any striking quotes that support the key logics.