动态异构特征融合的水下图像增强算法
作者:
作者单位:

沈阳大学

作者简介:

通讯作者:

中图分类号:

TP391

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),


Underwater Image Enhancement Algorithm Based on Dynamic Heterogeneous Feature Fusion
Author:
Affiliation:

Shenyang University

Fund Project:

The National Natural Science Foundation of China (General Program, Key Program, Major Research Plan)

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

    针对水下图像细节模糊以及色彩失真严重的问题,本文提出一种基于编码解码结构的动态异构特征融合水下图像增强网络。首先,设计异构特征融合模块,将不同级别与不同层次的特征进行融合,提升网络对细节信息与语义信息的整体感知能力。其次,设计新型特征注意力机制,改进传统通道注意力机制,并将改进后的通道注意力与像素注意力机制加入到异构特征融合过程,加强网络提取不同浑浊度像素特征的能力。然后,设计动态特征增强模块,自适应扩展感受野以提升网络对图像畸变景物的适应力及模型转换能力,加强网络对感兴趣区域的学习。最后,设计色彩损失函数,并联合最小化绝对误差损失与结构相似性损失,在保持图像纹理的基础上纠正色偏。实验结果表明,本文算法可以有效提升网络的特征提取能力,降低水下图像的雾度效应,提升图像的清晰度及色彩饱和度。

    Abstract:

    Aiming at the problems of blurred details of underwater images and serious color distortion, this paper proposes a dynamic heterogeneous feature fusion underwater image enhancement network based on the autoencoder structure. First, design a heterogeneous feature fusion module to integrate different levels and different levels of features to improve the overall perception of detailed information and semantic information of the network. Second, design a new feature attention mechanism, improve the traditional channel attention mechanism, and add the improved channel attention and pixel attention mechanism to the heterogeneous feature fusion process to strengthen the network"s ability to extract pixel features of different turbidity. Then, a dynamic feature enhancement module is designed to adaptively expand the receptive field to improve the network"s adaptability to image distortion scenes and model conversion capabilities, and strengthen the network"s learning of regions of interest. Finally, design the color loss function, and jointly minimize the absolute error loss and the structural similarity loss, and correct the color cast on the basis of maintaining the image texture. The experimental results show that the algorithm in this paper can effectively improve the feature extraction ability of the network, reduce the haze effect of underwater images, and improve the clarity and color saturation of the image.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-10-25
  • 最后修改日期:2022-09-03
  • 录用日期:2022-03-15
  • 在线发布日期: 2022-04-01
  • 出版日期: