Research of Different Neural Network Architectures for Audio and Video Denoising
Anton I. Kanev, М. В. Назаров, Daniel Uskov, Vladislav I. Terentyev
Abstract
Recently, neural networks have attracted considerable attention due to their better denoising quality compared to the previously used frequency-time filters for audio denoising and algorithmic methods for visual denoising. In this paper different neural network architectures used for denoising from audio and video files are discussed. The architectures of two neural networks RNNoise and PoCoNet are compared to audio denoising. The main source of visual noise in video is Gaussian noise, which occurs during digital imaging of real objects, such as sensor noise caused by poor light or high temperature, and electronic noise. Two solutions to neural network architectures are compared to remove Gaussian noise from video. In the first approach, each frame is processed as a separate image. For this approach, the DnCNN and Restormer neural networks have been considered. In the second approach, frames are considered as a sequence of images over time. In this approach, FastDVDnet and PaCNet neural networks have been considered. The aim is to compare different neural networks architectures for removing audio and video noise from public room camera recordings.