Exploring Efficient and Tunable Convolutional Blind Image Denoising Networks

Abstract

We address the problem of building a blind image denoising network that better adapts to user-defined efficiency and performance requirements. CNN-based architectures such as FFDNet as well as classical methods like BM3D provide fast denoising capability but require the user to specify an approximate noise level. Blind denoising networks like DnCNN and CBDNet are appealing due to their ease of use by non-experts but can be slow. Additionally, these networks are not designed to allow for selecting a reliable operating point based on constraints like available compute, affordable latency, and expected quality. To this end, we propose to develop denoising networks that are tunable to achieve a desired balance between image quality and model size. We seek inspiration from architectures that are tuned for classification, detection, and semantic segmentation on mobile phone CPUs. Incorporating recent advances in architectural building blocks and network architecture search and building upon the success of the DnCNN architectures, we present an efficient convolutional blind image denoising network.