•  
  •  
 

Abstract

Targeting issues related to low contrast, blurring, and loss of detail prevalent in underwater image enhancement algorithms, we propose a dual-color space multiscale residual network (DMR-SCNet) based on SCNet. First, we introduce the HSV color space feature extraction module, which aims to optimize the color representation and saturation of underwater images. Subsequently, we propose the RGB color space denoising module, which focuses on repairing the content and structure of underwater images to enhance their clarity and visual quality. Finally, by designing the residual attention (RAB) module, we aim to further refine the detailed representation and feature extraction of underwater images.

The results obtained from assessments conducted on the UIEBD and EUVP datasets indicate that the proposed method outperforms current prevalent deep learning methodologies, showing superior performance in terms of the peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and underwater image quality measure (UIQM). Our findings moreover indicate that DMR-SCNet yields substantial improvements in underwater image enhancement within diverse underwater environments. This approach shows promise for implementation in underwater image processing to markedly enhance the overall quality and usefulness of underwater imagery, offering broad applications in underwater visual data enhancement.

Download
COinS