A temporally-aware noise-informed invertible network for progressive video denoising

Abstract

Video denoising is a critical task in computer vision, aiming to enhance video quality by removing noise from consecutive video frames. Despite significant progress, existing video denoising methods still suffer from challenges in maintaining temporal consistency and adapting to different noise levels. To address these issues, a temporally-aware and noise-informed invertible network is proposed by following divide-and-conquer principle for progressive video denoising. Specifically, a recurrent attention-based reversible network is designed to distinctly extract temporal information from consecutive frames, thus tackling the learning problem of temporal consistency. Simultaneously, a noise-informed two-way dense block is developed by using estimated noise as conditional guidance to adapt to different noise levels. The noise-informed guidance can then be used to guide the learning of dense block for efficient video denoising. Under the framework of invertible network, the designed two parts can be further integrated to achieve invertible learning to enable progressive video denoising. Experiments and comparative studies demonstrate that our method can achieve good denoising accuracy and fast inference speed in both synthetic scenes and real-world applications.