+高级检索
基于小波变换及注意力机制的T型图像去雾网络
作者:
作者单位:

兰州交通大学

作者简介:

通讯作者:

基金项目:

国家自然科学基金项目(面上项目,重点项目,重大项目)


T-shaped image dehazing network based on wavelet transform and attention mechanism
Author:
Affiliation:

Lanzhou Jiao Tong University

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
    摘要:

    受大气中雾霾等悬浮颗粒的影响,室外拍摄的图像常伴有低对比度和低能见度的问题,现存去雾方法未能充分利用图像的局部特征信息,且不能完整提取图像的全局细节特征,因此存在有去雾不彻底及图像细节丢失等问题。为此,本文提出了一种基于小波变换及注意力机制的T型图像去雾网络。具体来说,所提网络通过对图像进行多次离散小波分解及重构来获取有雾图像的边缘细节特征,并提出了一种兼顾图像全局特征及局部信息提取的特征注意力模块,加强了网络在图像视觉感知和细节纹理方面的学习。其次,在进行特征提取的过程中提出T型连接方式来获得多尺度的图像特征,扩展了网络的表示能力。最后,对重构后的无雾图像进行色彩平衡,得到最终复原图像。在合成数据集和真实数据集中的大量实验结果表明,本文所提网络相较于现有其他网络模型具有更优越的性能。

    Abstract:

    Affected by suspended particles such as haze in the atmosphere,images taken outdoors often suffer from low contrast and low visibility. Existing dehazing methods fail to make full use of the local feature information of the image, and cannot fully extract the global details of the image. Therefore, there are problems such as incomplete dehazing and loss of image details. For this reason, this paper proposes a T-shaped image dehazing network based on wavelet transform and attention mechanism. Specifically, the proposed network obtains the edge detail features of the hazy image by performing multiple discrete wavelet decomposition and reconstruction on the image, and proposes a feature attention module that takes into account both the global feature and the local information extraction of the image, which strengthens the network"s learning in image visual perception and detail texture. Secondly, in the process of feature extraction, a T-shaped method is proposed to obtain multi-scale image features, which expands the network"s representation ability. Finally, perform color balance on the reconstructed clear image to obtain the final restored image. A large number of experimental results in synthetic data sets and real data sets show that the network proposed in this paper has superior performance compared with other existing network models.

    参考文献
    相似文献
    引证文献
文章指标
  • PDF下载次数:
  • HTML阅读次数:
  • 摘要点击次数:
  • 引用次数:
引用本文
历史
  • 收稿日期: 2021-10-02
  • 最后修改日期: 2021-10-27
  • 录用日期: 2021-10-28
  • 在线发布日期: 2022-05-17
  • 出版日期:
版权声明:稿件一经被本刊录用,即视为作者将版权授予本刊,本刊有权根据工作需要,允许合作的数据库、新媒体平台及其他媒体进行二次转载、推介、下载和传播,如有异议,请在来稿中声明。
关闭