|M.Sc Student||Mauda Assaf|
|Subject||Poisson Denoising of Images using Deep Neural Networks|
Inspired by Classical Dictionary based Algorithms
|Department||Department of Computer Science||Supervisor||PROF. Michael Elad|
Removing Poisson noise from low-light images is an important and fundamental task in image processing. While most of the research on image denoising focus on Gaussian noise, Poisson denoising is essential for many applications, including night vision, astronomy and fluorescence microscopy. Several approaches exist for this challenging task, some handle the Poisson noise distribution directly, some leverage a Gaussian denoiser and newer deep learning approaches often ignore the specific noise distribution when selecting their model, relying only on a supervised learning strategy.
Leveraging a Gaussian denoiser is mostly done using a variance stabilizing transformation (VST), which transforms a Poisson noisy image into an image with an approximately Gaussian noise. A Gaussian denoiser can then be used, and an inversed transform yields the clean image. While this method is widely used and works well for higher SNR image signals, the Gaussian approximation fails for very low photon-counts, and a different approach should be considered. For example, using the Plug-and-Play Prior method, the Poisson denoising of an image can be performed in an iterative manner using a Gaussian denoiser across the entire image intensity range. Deep learning methods use a large dataset of images to train parametric denoisers and achieve state-of-the-art results. Nevertheless, the interpretability and explainability of these denoisers is lacking. To combat these problems, a possible approach is to adapt classical image processing principles and algorithms into the denoiser architecture.
In this work we propose several novel deep learning networks for Poisson denoising,
all relying on classical image processing principles. The proposed solutions are based on sparse representation and multiscale analysis principles, and their action and learned parameters can be explained theoretically. We compare the performance of these models to other common classical and deep learning algorithms and show the competitive results of these models.