基于循环神经网络的多阶段图像去噪方法  

Multi-stage Image Denoising Based on Recurrent Neural Network

作　　者：林煌伟 陈钧荣 牛玉贞[1,2] LIN Huang-wei;CHEN Jun-rong;NIU Yu-zhen(College of Mathematics and Computer Science,Fuzhou University,Fuzhou 350105,China;Key Laboratory of Spatial Data Mining&Information Sharing,Ministry of Education,Fuzhou 350105,China)

机构地区：[1]福州大学数学与计算机科学学院,福州350105 [2]空间数据挖掘与信息共享省部共建教育部重点实验室,福州350105

出　　处：《小型微型计算机系统》2022年第1期84-90,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61672158)资助;福建省自然科学基金重点项目(2019J02006)资助;福建省自然科学基金面上项目(2020J001494)资助。

摘　　要：图像在采集和传输过程中往往受到噪声污染,去噪任务是图像预处理中的重要步骤.现有的基于深度学习的图像去噪方法往往只进行单次的去噪,容易产生过于平滑或者太多噪点未去除的结果且无法恢复.因此,本文提出了一种基于循环神经网络的多阶段图像去噪方法.该方法包括两个去噪阶段,通过调整两个阶段的训练权重可以使得第1阶段的去噪结果包含部分未去除干净的噪点和更多细节信息,然后将第1阶段提取的特征通过门控循环单元传递到第2阶段,再进行第2个阶段的去噪.同时,为了使深度网络的训练更稳定,本文还设计了一个估计噪声分布的子网络,用于从噪声图像中估计噪声的分布.最后,将噪声分布和噪声图像拼接,作为网络的输入来训练去噪网络.实验结果表明,本文的基于循环神经网络的多阶段图像去噪方法具有先进的去噪性能.Images are often contaminated by noise during the process of acquisition and transmission, the task of denoising is an important step in image preprocessing.Existing image denoising methods based on deep learning often perform denoising operation only once, it is easy to cause detail loss in denoising results.Therefore, this paper presents a multi-stage image denoising method based on recurrent neural network.This method includes two denoising stages.By adjusting the training weights, the denoising result of first stage will contain some unremoved noise and remain much image details, and the features extracted in first stage will pass to the second stage through Gated Recurrent Unit, and then the second stage of denoising operation is carried out.At the same time, in order to make the training of the deep network more stable, we also designed a sub-network to estimate the noise distribution, which is used to reconstruct noise distribution from the original noisy image.Finally, we concatenate the noise distribution image and the noisy image as the input of the network to train the denoising network.Experimental results show that proposed multi-stage image denoising method based on recurrent neural network achieves advanced denoising performance.

关 键 词：图像去噪 循环神经网络 多阶段去噪 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]