欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2021, Vol. 42 ›› Issue (6): 948-956.DOI: 10.11996/JG.j.2095-302X.2021060948

• 图像处理与计算机视觉 • 上一篇    下一篇

门控融合对抗网络的水下图像增强 

  

  1. 1. 沈阳理工大学自动化与电气工程学院，辽宁 沈阳 110159；  2. 辽宁工程技术大学电子与信息工程学院，辽宁 葫芦岛 125105
  • 出版日期:2022-01-18 发布日期:2022-01-18
  • 基金资助:
    国家重点研发计划项目(2018YFB1403303)；辽宁省重点研发计划资助项目(2019JH2/10100014) 

Underwater image enhancement algorithm using gated fusion generative adversarial network 

  1. 1. School of Automation and Electrical Engineering, Shenyang Ligong University, Shenyang Liaoning 110159, China;  2. School of Electronic and Information Engineering, Liaoning Technical University, Huludao Liaoning 125105, China
  • Online:2022-01-18 Published:2022-01-18
  • Supported by:
    National Research Key Program of China (2018YFB1403303); Liaoning Provincial Research Foundation for Key Program of China (2019JH2/10100014)

PDF (PC)

103

摘要： 针对水下成像中图像存在的色彩失真、对比度低和雾化严重等问题，提出一种门控融合对抗网 络的水下图像增强方法。主要特点是使用生成网络将图像特征信息逐像素恢复，并通过门控融合形成清晰图像。 首先，利用多个并行子网络对同幅图像的不同空间特征信息进行学习，提升网络对图像特征学习的多样性。然 后，通过门控融合，将不同子网络学习到的图像特征相融合。利用生成网络与鉴别网络进行相互博弈，反复训 练网络，获得增强的水下图像。最后，在 EUVP 数据集和 U45 测试集上进行实验对比。实验结果表明，该算 法的关键点匹配与原图相比平均高 19 个匹配点，UCIQE 平均值为 0.664 7，UIQM 平均值为 5.723 7，与其他经 典及最新算法相比具有优势，效果良好。

关键词: 图像处理, 水下图像, 生成对抗网络, 多尺度输入, 门控融合

Abstract: An underwater image enhancement algorithm using gated fusion generative adversarial network was proposed for solving problems of underwater image color distortion, low contrast, and heavy fogging. The crucial element of this algorithm is that it recruited a generator to pixel-by-pixel restore image feature details and synthesized a clear image through gated fusion. First of all, to increase the variety of image feature learning by the network, several parallel sub-networks were employed to learn different kinds of spatial feature knowledge of the same image. The image features learned from different sub-networks were fused utilizing gated fusion. The generator and the discriminator were used for mutual games, and the network was repeatedly trained to obtain enhanced underwater images. Finally, using the EUVP dataset and the U45 testset, this paper performed a series of comparative experiments. The algorithm’s key point matching was 19 points higher than the raw image, according to the experimental results. The average UCIQE value was 0.664 7, while the average UIQM value was 5.723 7. It can achieve improvements over other classic and latest algorithms, demonstrating the algorithm’s extraordinary performance. 

Key words: image processing, underwater image, generative adversarial network, multi-scale input, gated fusion

中图分类号: