Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (24): 196-208.DOI: 10.3778/j.issn.1002-8331.2207-0146

• Graphics and Image Processing • Previous Articles     Next Articles

Research on Multi-Branch Image Denoising Algorithm

GENG Jun, LI Wenhai, WU Zihao, SUN Xinjie   

  1. College of Software, Xinjiang University, Urumqi 830091, China
  • Online:2023-12-15 Published:2023-12-15

多分支图像去噪算法研究

耿俊,李文海,吴子豪,孙鑫杰   

  1. 新疆大学 软件学院,乌鲁木齐 830091

Abstract: In recent years, deep convolutional neural network(CNN) has caused a great sensation in the field of image denoising. However, for the Gaussian image denoising task, it has some problems:(1) most of the single-branch models can not make full use of image features and are often affected by information loss. (2) Most deep CNNs have the problem of insufficient edge feature extraction and performance saturation. In order to solve these two problems, a multi-branch network model based on deep learning (MBNet) is proposed. Firstly, in order to solve the problem of insufficient feature extraction of single-branch network model, MBNet introduces multiple different and complementary networks to combine and then perform feature fusion to enhance the denoising effect and generalization ability. Secondly, in order to solve the problem of inadequate edge feature extraction, MBNet introduces multiple cavity convolution with different expansion rates to increase the receptive field and extract more context information. Finally, in order to solve the performance saturation problem of deep CNN, MBNet also adopts multi-local residual learning and global residual learning. A large number of experimental results show that when [σ=15], the average PSNR values of MBNet on Set12, BSD68, CBSD68, Kodak24 and McMaster datasets are 32.981 dB, 31.750 dB, 34.001 dB, 34.709 dB and 34.394 dB, respectively. MBNet has better performance than the current advanced image denoising methods, and can obtain clearer image and edge texture features in subjective visual effects.

Key words: convolutional neural network, residual learning, batch normalization, dilated convolution

摘要: 近些年,深卷积神经网络(CNN)在图像去噪领域引起了极大的轰动。然而,对于高斯图像去噪任务来说,它有一些问题:(1)绝大多数的单分支模型不能充分地利用图像特征,经常受到信息丢失的影响;(2)大多数深度CNN存在边缘特征提取不足且性能饱和问题。为了解决这两个问题,提出了基于深度学习的多分支网络模型(multi-branch network model based on deep learning,MBNet)。为了解决单分支网络模型提取特征不充分问题,MBNet引入多个不同且互补的网络相结合再进行特征融合来增强去噪效果和泛化能力;为了解决边缘特征提取不充分问题,MBNet引入多个不同扩张率的空洞卷积来增大感受野,提取更多的上下文信息;为了解决深度CNN性能饱和问题,MBNet还采用了多局部残差学习和整体残差学习的方式。大量实验结果表明,当[σ=15]时,MBNet在Set12、BSD68、CBSD68、Kodak24、McMaster数据集上的平均PSNR值分别为32.981?dB、31.750?dB、34.001?dB、34.709?dB、34.394?dB。MBNet比目前先进的图像去噪方法具有更好的性能,并且在主观视觉效果上得到更加清晰的图像和边缘纹理特征。

关键词: 卷积神经网络, 残差学习, 批量归一化, 空洞卷积