摘要
现有卷积神经网络去噪算法大多只能在合成噪声上表现良好,且只从单一尺度上提取特征,无法构建更干净的图像。针对以上问题,提出了一种基于卷积神经网络的多尺度特征融合去噪算法,该算法利用分层得到不同尺度的特征,可以获得更多的感受野,通过特征融合将上一尺度的特征融合到当前尺度中,进行渐进式的训练,可以去除更多的噪声。在每一层的特征提取后设置一个编-解码器,并在编-解码器中加入空洞卷积,防止图像分辨率过低导致图像细节破坏、信息丢失。用提出的网络模型在合成噪声数据集和真实噪声数据集上进行实验,结果表明,该算法的去噪性能优于对比算法,能够保留更多的细节。
Most of the existing convolutional neural network denoising algorithms can only perform well in synthetic noise,and only extract features from a single scale,which can not reconstruct a cleaner image.To solve the above problems,a multi-scale feature fusion denoising algorithm based on convolution neural network is proposed.The algorithm uses layered features of different scales to obtain more receptive fields.Through feature fusion,the features of the previous scale are fused to the current scale for progressive training,which can remove more noise.A encode-decode is set after the feature extraction of each layer,and hole convolution is added to the encode-decode to prevent image detail damage and information loss when the image resolution is too low.The proposed network model is tested on synthetic noise data set and real noise data set.The results show that the denoising performance of the algorithm is better than the comparison algorithms and can retain more details.
作者
许雪
郭业才
李晨
XU Xue;GUO Yecai;LI Chen(College of Electronic and Information Engineering,Nanjing University of Information Science and Technology,Nanjing 210044;College of Electronic Information,Wuxi University,Wuxi 214105)
出处
《计算机与数字工程》
2023年第10期2400-2404,2417,共6页
Computer & Digital Engineering
基金
国家自然科学基金项目(编号:61673222)
无锡学院人才启动经费项目(编号:550221028)资助。
关键词
图像处理
卷积神经网络
多尺度特征
特征融合
真实噪声图像去噪
image processing
convolutional neural network
multiscale features
feature fusion
real-world noisy image denoising
作者简介
许雪,女,硕士,研究方向:深度学习与图像去噪;郭业才,男,博士,教授,博士生导师,研究方向:通信信号处理、气象信息技术与安全、水声信号处理等;李晨,女,博士,讲师,研究方向:深度学习与图像复原。
