MANet: Improving Video Denoising with a Multi-Alignment Network

TL;DR

This paper introduces MANet, a multi-alignment network for video denoising that uses multiple flow proposals and attention to improve noise suppression, outperforming baseline models and reducing parameters.

Contribution

The paper proposes a novel multi-alignment network that enhances video denoising by mimicking non-local mechanisms through multiple flow proposals and attention-based averaging.

Findings

Improves denoising performance by 0.2dB over baseline.

Reduces model parameters by 47% with distillation.

Applicable to various flow-based denoising models.

Abstract

In video denoising, the adjacent frames often provide very useful information, but accurate alignment is needed before such information can be harnassed. In this work, we present a multi-alignment network, which generates multiple flow proposals followed by attention-based averaging. It serves to mimic the non-local mechanism, suppressing noise by averaging multiple observations. Our approach can be applied to various state-of-the-art models that are based on flow estimation. Experiments on a large-scale video dataset demonstrate that our method improves the denoising baseline model by 0.2dB, and further reduces the parameters by 47% with model distillation. Code is available at https://github.com/IndigoPurple/MANet.