Helou Majed El ; Süsstrunk Sabine

Abstract

Blind and universal image denoising consists of using a unique model thatdenoises images with any level of noise. It is especially practical as noiselevels do not need to be known when the model is developed or at test time. Wepropose a theoretically-grounded blind and universal deep learning imagedenoiser for additive Gaussian noise removal. Our network is based on anoptimal denoising solution, which we call fusion denoising. It is derivedtheoretically with a Gaussian image prior assumption. Synthetic experimentsshow our network's generalization strength to unseen additive noise levels. Wealso adapt the fusion denoising network architecture for image denoising onreal images. Our approach improves real-world grayscale additive imagedenoising PSNR results for training noise levels and further on noise levelsnot seen during training. It also improves state-of-the-art color imagedenoising performance on every single noise level, by an average of 0.1dB,whether trained on or not.