Variational Bayesian deep network for blind Poisson denoising

Abstract

Deep learning-based approaches have recently achieved considerable results in Poisson denoising under low-light conditions. However, most existing methods mainly focus on the network architecture design, which lacks physical interpretability and thus unsuitable for blind denoising in real environments with unknown levels of noises. To address this issue, we propose a variational Bayesian deep network for blind Poisson denoising (VBDNet). We mainly consider an approximate posterior form for the noise variance in a variational Bayesian framework and utilize a neural network to parameterize the variance of Poisson noise. For network design, VBDNet is divided into two sub-networks. The noise estimation sub-network is responsible for the Bayesian inference. This network improves the blind denoising ability of the subsequent denoising sub-network by learning Poisson noise characteristics under different noise levels in the training process. A network of U-Net structures implements the denoising sub-network for noise removal. By combining the advantage of Bayesian inference (noise estimation sub-network) and deep learning (denoising sub-network), VBDNet outperforms other state-of-the-art methods on both synthetic and natural data. The code and details are available at https://github.com/HLImg/VBDNet.