Deep Iterative Down-Up CNN for Image Denoising

Networks using down-scaling and up-scaling of feature maps have been studied extensively in low-level vision research owing to efficient GPU memory usage and their capacity to yield large receptive fields. In this paper, we propose a deep iterative down-up convolutional neural network (DIDN) for image denoising, which repeatedly decreases and increases the resolution of the feature maps. The basic structure of the network is inspired by U-Net which was originally developed for semantic segmentation. We modify the down-scaling and up-scaling layers for image denoising task. Conventional denoising networks are trained to work with a single-level noise, or alternatively use noise information as inputs to address multi-level noise with a single model. Conversely, because the efficient memory usage of our network enables it to handle multiple parameters, it is capable of processing a wide range of noise levels with a single model without requiring noise-information inputs as a work-around. Consequently, our DIDN exhibits state-of-the-art performance using the benchmark dataset and also demonstrates its superiority in the NTIRE 2019 real image denoising challenge.