Scholarlink

<p>Underwater image enhancement (UIE) aims to solve or mitigate issues such as color cast, low contrast, and blurred details in underwater images, having wide applications in fields such as ocean exploration, marine resource development, underwater archaeology, and underwater robotics. However, many UIE methods still fail to adequately address the issue of poor domain generalization ability. One possible reason is that they lack the integration of image content and style control. Additionally, they may generate low visual quality enhanced images due to unrealistic content and unattractive styles, especially methods based on Cycle-Consistent Generative Adversarial Network (CycleGAN). To address these issues, we propose a Content-Style Control Network with Style Contrastive Learning (CSC-SCL). Firstly, to enhance the domain generalization ability of the model in handling underwater images with different distortion distributions, we design a generator to extract domain-invariant features by adaptively adjusting the degree of content and style changes at the feature level without requiring domain labels. Then, to enhance visual quality in terms of both content and style, we design content loss and style contrastive learning, which are used to ensure the authenticity of the image content and produce more pleasing styles, respectively. Extensive experiments on three datasets demonstrate that our method outperforms most existing methods in terms of both visual quality and quantitative metrics. Our code will be released at <a href="https://github.com/xjq-flare/CSC-SCL">https://github.com/xjq-flare/CSC-SCL</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

A content-style control network with style contrastive learning for underwater image enhancement

  • 期刊论文

  • 发表日期：

    2025年1月9日

查看全文
  • Zhenguang Wang,
  • Huanjie Tao,
  • Hui Zhou,
  • Yishi Deng,
  • Ping Zhou

摘要

Underwater image enhancement (UIE) aims to solve or mitigate issues such as color cast, low contrast, and blurred details in underwater images, having wide applications in fields such as ocean exploration, marine resource development, underwater archaeology, and underwater robotics. However, many UIE methods still fail to adequately address the issue of poor domain generalization ability. One possible reason is that they lack the integration of image content and style control. Additionally, they may generate low visual quality enhanced images due to unrealistic content and unattractive styles, especially methods based on Cycle-Consistent Generative Adversarial Network (CycleGAN). To address these issues, we propose a Content-Style Control Network with Style Contrastive Learning (CSC-SCL). Firstly, to enhance the domain generalization ability of the model in handling underwater images with different distortion distributions, we design a generator to extract domain-invariant features by adaptively adjusting the degree of content and style changes at the feature level without requiring domain labels. Then, to enhance visual quality in terms of both content and style, we design content loss and style contrastive learning, which are used to ensure the authenticity of the image content and produce more pleasing styles, respectively. Extensive experiments on three datasets demonstrate that our method outperforms most existing methods in terms of both visual quality and quantitative metrics. Our code will be released at https://github.com/xjq-flare/CSC-SCL.