计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (18): 179-189.DOI: 10.3778/j.issn.1002-8331.2205-0204

• 图形图像处理 • 上一篇    下一篇

加权密集扩张卷积网络的随机脉冲噪声去除

曹义亲,符杨逸,饶哲初   

  1. 华东交通大学 软件学院,南昌 330013
  • 出版日期:2023-09-15 发布日期:2023-09-15

Weighted Dense Dilated Convolutional Network for Random Impulse Noise Removal

CAO Yiqin, FU Yangyi, RAO Zhechu   

  1. School of Software, East China Jiaotong University, Nanchang 330013, China
  • Online:2023-09-15 Published:2023-09-15

摘要: 基于深度学习的图像去噪方法,大多没有充分利用不同层次的特征信息,通道合并时都是直接在通道维度上对特征图进行拼接,并没有考虑到浅层与深层卷积特征各自的重要性。为解决上述问题,提出一种加权密集扩张卷积连接网络模型,用于去除图像的随机脉冲噪声。通过使用不同扩张因子的扩张卷积来丰富浅层特征图的多尺度特征信息;考虑到浅层与深层卷积特征各自的重要性,将原始密集块进行改进,采用加权密集连接结构,并使用扩张卷积提高感受野;采用跳跃连接,将浅层的多尺度特征信息和不同加权密集扩张卷积块的特征信息进行融合,充分利用深层卷积特征和浅层卷积特征信息实现随机脉冲噪声的复原。实验结果表明,所提模型的去噪效果更加突出。

关键词: 图像去噪, 深度学习, 密集连接, 扩张卷积, 权重, 随机脉冲噪声

Abstract: Among image denoising methods based on deep learning, most of them do not make full use of feature information at different levels, and splicing feature graphs directly on channel dimensions during channel merging, without considering the importance of both shallow and deep convolution features. In order to solve the above problems, this paper proposes a weighted dense dilated convolutional connection network to remove random impulse noise in images. Firstly, the multi-scale feature information of the shallow feature map is enriched by dilation convolution using different expansion factors. Secondly, considering the importance of both shallow and deep convolution features, the original dense block is improved, the weighted dense connection structure is adopted, and the expansion convolution is used to effectively increase the receptive field. Finally, jump connections are used to fuse the shallow multi-scale feature information with the feature information of different weighted dense dilated convolution blocks, and make full use of the high-level and low-level feature information to recover random impulse noise. The experimental results show that the proposed algorithm has a more prominent denoising.

Key words: image denoising, deep learning, dense connection, dilated convolution, weight, random impulsive noise