Scholarlink

<p>Scattering and absorption in underwater environments deteriorate image quality by reducing visibility and distorting color balance. Thus, Underwater Image Enhancement (UIE) constitutes a crucial pre-processing step for facilitating high-level vision tasks in aquatic environments. However, existing methods often suffer from high computational costs and limited feature extraction. In response to these challenges, this study proposes GD-LiteNet, a lightweight Convolutional Neural Network (CNN) with Grouped Depthwise Convolution and Dynamic Attention for efficient UIE. The architecture integrates novel Grouped Depthwise Separable Convolution (GDSC) and Dynamic Squeeze-and-Excitation (DSE) modules. GDSC enhances feature extraction while significantly reducing computational complexity, whereas DSE adaptively recalibrates channel-wise attention to emphasize critical regions. Experimental evaluations on benchmark underwater datasets (UIEB and EUVP) show GD-LiteNet achieves a PSNR of 28.31, SSIM of 0.9119, MSE of 0.0009, and UIQM of 3.15. GD-LiteNet significantly reduces parameters to 8.768K and achieves an inference time of 0.0341 seconds, outperforming the most efficient baseline.</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

GD-LiteNet: underwater image enhancement using lightweight CNN with grouped depthwise convolution and dynamic attention

  • 期刊论文

  • 发表日期：

    2026年6月17日

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

摘要

Scattering and absorption in underwater environments deteriorate image quality by reducing visibility and distorting color balance. Thus, Underwater Image Enhancement (UIE) constitutes a crucial pre-processing step for facilitating high-level vision tasks in aquatic environments. However, existing methods often suffer from high computational costs and limited feature extraction. In response to these challenges, this study proposes GD-LiteNet, a lightweight Convolutional Neural Network (CNN) with Grouped Depthwise Convolution and Dynamic Attention for efficient UIE. The architecture integrates novel Grouped Depthwise Separable Convolution (GDSC) and Dynamic Squeeze-and-Excitation (DSE) modules. GDSC enhances feature extraction while significantly reducing computational complexity, whereas DSE adaptively recalibrates channel-wise attention to emphasize critical regions. Experimental evaluations on benchmark underwater datasets (UIEB and EUVP) show GD-LiteNet achieves a PSNR of 28.31, SSIM of 0.9119, MSE of 0.0009, and UIQM of 3.15. GD-LiteNet significantly reduces parameters to 8.768K and achieves an inference time of 0.0341 seconds, outperforming the most efficient baseline.