Scholarlink

<p>The underwater images directly obtained from the light vision system usually suffer from serious degradation due to the complex imaging environment. To improve the visual quality of original images, many physical models or deep learning based methods have been proposed. However, the real physical imaging process of underwater images is hard to formulate, and the training process of deep networks needs a lot of paired training samples, which are also hard to obtain. To address these issues, an image restoration method via domain transfer learning and physical aware deep networks has been proposed. Firstly, a domain transfer framework is devised based on the Retinex theory and feature enhanced encoder-decoder network. More accurate characterization of water is learned, hence the generated underwater images are closer to the real ones. Furthermore, the physical aware deep networks are proposed, which integrate the end-to-end deep neural networks and physical imaging process into a whole framework. The connection between clear in-air and blurred underwater domains has been established, thus the underwater image restoration result is closer to the clear in-air image. Besides, more detailed information could be restored by benefiting from the respective advantages of deep learning and the simulation of underwater image degradation. Qualitative and quantitative experimental results on both synthetic and real-world datasets demonstrate that the proposed method could achieve superior restoration results compared with other state-of-the-art approaches. The code is available at <a href="https://github.com/UIEDM/UIEDTRPADN">https://github.com/UIEDM/UIEDTRPADN</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Underwater image restoration via domain transfer learning and physical aware deep networks

  • 期刊论文

  • 发表日期：

    2026年3月23日

查看全文
  • Tingting Yao,
  • Zihao Feng,
  • Yuan Gao,
  • Qing Hu,
  • Na Xia

摘要

The underwater images directly obtained from the light vision system usually suffer from serious degradation due to the complex imaging environment. To improve the visual quality of original images, many physical models or deep learning based methods have been proposed. However, the real physical imaging process of underwater images is hard to formulate, and the training process of deep networks needs a lot of paired training samples, which are also hard to obtain. To address these issues, an image restoration method via domain transfer learning and physical aware deep networks has been proposed. Firstly, a domain transfer framework is devised based on the Retinex theory and feature enhanced encoder-decoder network. More accurate characterization of water is learned, hence the generated underwater images are closer to the real ones. Furthermore, the physical aware deep networks are proposed, which integrate the end-to-end deep neural networks and physical imaging process into a whole framework. The connection between clear in-air and blurred underwater domains has been established, thus the underwater image restoration result is closer to the clear in-air image. Besides, more detailed information could be restored by benefiting from the respective advantages of deep learning and the simulation of underwater image degradation. Qualitative and quantitative experimental results on both synthetic and real-world datasets demonstrate that the proposed method could achieve superior restoration results compared with other state-of-the-art approaches. The code is available at https://github.com/UIEDM/UIEDTRPADN.