Zhang Kai Zuo Wangmeng Zhang Lei

Abstract

Due to the fast inference and good performance, discriminative learningmethods have been widely studied in image denoising. However, these methodsmostly learn a specific model for each noise level, and require multiple modelsfor denoising images with different noise levels. They also lack flexibility todeal with spatially variant noise, limiting their applications in practicaldenoising. To address these issues, we present a fast and flexible denoisingconvolutional neural network, namely FFDNet, with a tunable noise level map asthe input. The proposed FFDNet works on downsampled sub-images, achieving agood trade-off between inference speed and denoising performance. In contrastto the existing discriminative denoisers, FFDNet enjoys several desirableproperties, including (i) the ability to handle a wide range of noise levels(i.e., [0, 75]) effectively with a single network, (ii) the ability to removespatially variant noise by specifying a non-uniform noise level map, and (iii)faster speed than benchmark BM3D even on CPU without sacrificing denoisingperformance. Extensive experiments on synthetic and real noisy images areconducted to evaluate FFDNet in comparison with state-of-the-art denoisers. Theresults show that FFDNet is effective and efficient, making it highlyattractive for practical denoising applications.