Scholarlink

<p>Underwater image enhancement is a critical challenge due to the unique optical properties of water, leading to issues such as low contrast, color distortion, and noise pollution. This paper proposes a lightweight generative adversarial network (GAN) architecture, MWCA-GAN, based on MobileNetV4, specifically designed for underwater image enhancement. The model incorporates a wavelet denoising module, a color correction module, and a multi-query attention mechanism (MQA) to address multi-source degradation in underwater images. A multi-scale discriminator is employed to ensure hierarchical quality supervision from global to local textures. The loss function combines Huber adversarial loss, Smooth L1 similarity loss, and content-aware loss based on VGG19 and ResNet50 to improve image structure and detail expression. Experimental results on the EUVP and UIEBD datasets demonstrate that MWCA-GAN achieves PSNR values of 27.65&#xa0;dB and 23.35&#xa0;dB, respectively, and SSIM values of 0.818 and 0.858, respectively, outperforming mainstream methods in terms of UIQM and UCIQE. This work not only advances the field of underwater image enhancement but also provides a lightweight solution suitable for resource-constrained devices.</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

Enhancing Underwater Image Quality: A Lightweight GAN Approach with MobileNetV4 and Multi-scale Discrimination

  • 期刊论文

  • 发表日期：

    2025年11月30日

查看全文
  • Zhou Zhiyu,
  • Shengxuan Lin,
  • Yufeng Qiu,
  • Laihu Peng,
  • Haiyan Wang

摘要

Underwater image enhancement is a critical challenge due to the unique optical properties of water, leading to issues such as low contrast, color distortion, and noise pollution. This paper proposes a lightweight generative adversarial network (GAN) architecture, MWCA-GAN, based on MobileNetV4, specifically designed for underwater image enhancement. The model incorporates a wavelet denoising module, a color correction module, and a multi-query attention mechanism (MQA) to address multi-source degradation in underwater images. A multi-scale discriminator is employed to ensure hierarchical quality supervision from global to local textures. The loss function combines Huber adversarial loss, Smooth L1 similarity loss, and content-aware loss based on VGG19 and ResNet50 to improve image structure and detail expression. Experimental results on the EUVP and UIEBD datasets demonstrate that MWCA-GAN achieves PSNR values of 27.65 dB and 23.35 dB, respectively, and SSIM values of 0.818 and 0.858, respectively, outperforming mainstream methods in terms of UIQM and UCIQE. This work not only advances the field of underwater image enhancement but also provides a lightweight solution suitable for resource-constrained devices.