Scholarlink

<p>Enhancing underwater images is vital for improving visibility and supporting applications in marine research, ocean archaeology, and surveillance. Underwater images are affected by color distortion, reduced contrast, and hazy appearances due to the way water absorbs and scatters light. To tackle these challenges, we present the proposed model, which is built on an encoder–decoder backbone and enriched with two key modules: Spectral–Instance Residual Block (SIRB) and Convolutional Block Attention Module (CBAM). Encoder extracts features at multiple scales from degraded inputs, and the decoder gradually reconstructs clear images by combining global context with fine details through skip connections. SIRB enhances feature stability and representation quality, whereas CBAM guides the proposed model to concentrate on the most relevant feature channels and spatial regions. A PatchGAN-inspired discriminator further strengthens the process by pushing the generator to produce outputs with realistic textures and natural color tones. We evaluate the proposed model on standard benchmarks including UIEB, EUVP, and U45, where it consistently outperforms existing methods in both quantitative metrics (PSNR, SSIM, MSE) and visual quality. Beyond benchmarks, we also tested the model on real tasks like edge detection, feature extraction, and image segmentation. The enhanced images gave more reliable results in these tasks, which shows the wider usefulness of the proposed model.</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 Enhancement with Spectral–Instance Residual and Attention-Guided Generative Networks

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Vaibhav Ingale,
  • Manish Kumar,
  • Dipika Gupta

摘要

Enhancing underwater images is vital for improving visibility and supporting applications in marine research, ocean archaeology, and surveillance. Underwater images are affected by color distortion, reduced contrast, and hazy appearances due to the way water absorbs and scatters light. To tackle these challenges, we present the proposed model, which is built on an encoder–decoder backbone and enriched with two key modules: Spectral–Instance Residual Block (SIRB) and Convolutional Block Attention Module (CBAM). Encoder extracts features at multiple scales from degraded inputs, and the decoder gradually reconstructs clear images by combining global context with fine details through skip connections. SIRB enhances feature stability and representation quality, whereas CBAM guides the proposed model to concentrate on the most relevant feature channels and spatial regions. A PatchGAN-inspired discriminator further strengthens the process by pushing the generator to produce outputs with realistic textures and natural color tones. We evaluate the proposed model on standard benchmarks including UIEB, EUVP, and U45, where it consistently outperforms existing methods in both quantitative metrics (PSNR, SSIM, MSE) and visual quality. Beyond benchmarks, we also tested the model on real tasks like edge detection, feature extraction, and image segmentation. The enhanced images gave more reliable results in these tasks, which shows the wider usefulness of the proposed model.