High Performance Video Quality Denoising by Deep Learning

The deep learning based video denoising method without motion estimation. We first stack the patches at the same spatial location across adjacent frames as a patch group, and learn a patch group prior model from a set of clean training videos. The patch group prior exploits the local temporal redundancy of video to remove noise.

A deep convolutional neural network is learned to exploit video spatial redundancy, as well as global temporal redundancy of video. The proposed algorithm exhibits better visual quality as well as quantitative measure than state-of-the-art video denoising methods and can be largely accelerated with GPU, provides a powerful tool for practical video denoising.