Image Denoising with Diversity Enhanced Conditional GAN

Image denoising is an inverse problem with many possible solutions. There have been many attempts to solve this problem, some producing single output, some a distribution of possible images. Among others, methods based on diffusion models and Generative Adversarial Networks (GANs) generate high-quality denoised images. GANs are particularly useful in high throughput applications due to their computational efficiency. Still, they often suffer from mode collapse and cannot provide diverse distributions of denoised images. In this work, we present a simple variance-enhanced conditional GAN denoising model. Training with additional mean stabilizing and variance-enhancing loss terms results in a model which outperforms standard methods on three simple data sets and is robust to the hyperparameter tuning. We show that this architecture is well suited for (high-throughput) setups where only noisy target images are present e.g. single particle electron microscopy (EM) image denoising. Finally, we show that in the case of noisy targets, best-denoised predictions can be selected by posterior sampling to enhance further processing steps.