FRDVDnet: Towards Fast and Robust Deep Video Denoising

With the rapid development of the technology of deep learning, lots of deep video denoising networks emerged. The mainstream deep video denoising(DVD) algorithms with the property of excellent detail preservation and the ability to process a wide range of noise level, such as FastDVDnet, are only based on the fixed datasets without enough generalizability. In order to deal with this problem, a network for fast and robust DVD called FRDVDnet, is proposed in this paper for optimizing FastDVDnet by utilizing dense-scale feature fusion mechanism and CReLU activation scheme. We compare FRDVDnet with FastDVDnet tested in the mid-scale bayer benchmark dataset visually and quantitatively. The simulation result shows that the denoising performance of FRDVDnet is improved with the increasing of the noise intensity compared to FastDVDnet. Moreover, in terms of perservation of key detail, FRDVDnet is superior to FastDVDnet on the denoised video.