Scholarlink

<p>High-quality underwater images are crucial for applications such as seafloor exploration, resource monitoring, and object detection. However, underwater imaging faces significant challenges, including poor visibility, low contrast, and severe color distortion. To address these issues, we propose a novel deep learning framework for Underwater Image Enhancement (UIE), incorporating color correction, cross-attention mechanism, and visibility enhancement. The proposed <b>M</b>ulti-<b>S</b>cale <b>C</b>olor <b>C</b>orrected <b>R</b>etinex <b>N</b>etwork (MSCC-RetNet) consists of three key modules: a multi-scale color correction module, an information complementarity module, and a multi-scale feature fusion module. The multi-scale color correction module integrates the Inception (Inc) and Squeeze-and-Excitation (SE) modules to enhance feature extraction. The multi-scale feature fusion module combines Convolutional Neural Networks (CNNs) with Transformers to improve representation learning. Experimental results indicate that the multi-scale color correction module effectively mitigates color distortions, while the cross-attention mechanism in the Information Complementarity module selectively emphasizes important regions. Additionally, the multi-scale fusion module enables end-to-end processing of illumination variations, significantly enhancing image visibility. Comprehensive evaluations on five benchmark underwater image datasets confirm that MSCC-RetNet outperforms four traditional approaches and five deep learning-based methods in key image quality assessment metrics.</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

MSCC-RetNet: a multi-scale color corrected retinex network for underwater image enhancement

  • 期刊论文

  • 发表日期：

    2025年5月14日

查看全文
  • Benxue Sun,
  • Mingxuan Chen,
  • Liming Hu,
  • Anjie Wang,
  • Zhijun Fang

摘要

High-quality underwater images are crucial for applications such as seafloor exploration, resource monitoring, and object detection. However, underwater imaging faces significant challenges, including poor visibility, low contrast, and severe color distortion. To address these issues, we propose a novel deep learning framework for Underwater Image Enhancement (UIE), incorporating color correction, cross-attention mechanism, and visibility enhancement. The proposed Multi-Scale Color Corrected Retinex Network (MSCC-RetNet) consists of three key modules: a multi-scale color correction module, an information complementarity module, and a multi-scale feature fusion module. The multi-scale color correction module integrates the Inception (Inc) and Squeeze-and-Excitation (SE) modules to enhance feature extraction. The multi-scale feature fusion module combines Convolutional Neural Networks (CNNs) with Transformers to improve representation learning. Experimental results indicate that the multi-scale color correction module effectively mitigates color distortions, while the cross-attention mechanism in the Information Complementarity module selectively emphasizes important regions. Additionally, the multi-scale fusion module enables end-to-end processing of illumination variations, significantly enhancing image visibility. Comprehensive evaluations on five benchmark underwater image datasets confirm that MSCC-RetNet outperforms four traditional approaches and five deep learning-based methods in key image quality assessment metrics.