Underwater Image Enhancement Using Channel-adaptive Frequency-spatial Network (CAFS-Net)

Deep learning (DL) has appeared as a powerful tool for addressing the limitations of conventional underwater image enhancement (UIE) techniques, which often struggle with inconsistent improvements and trade-offs between image attributes.This paper introduces CAFS-Net, a novel UIE network designed to tackle these challenges through a two-component approach: Frequency-Spatial Processing Network (FSPN) and Adaptive Channel Prior Modules (ACPM).FSPN systematically decouples and enhances individual degradation factors in UIs, leading to more balanced improvements without sacrificing one attribute for another.ACPM, combining Transformer and CNN (Convolutional Neural Network) models, effectively restores color by linking it to multi-scale image details using cross-attention.Our method operates in three stages: FSPN first enhances the image, ACPM then strengthens the relationship between image regions and color, and finally, IF-Net refines the overall details and colors.Experimental results demonstrate that CAFS-Net outperforms existing methods in both visual quality and quantitative metrics, achieving superior PSNR, SSIM, and UIQM scores, making it a robust solution for enhancing UIs in complex environments.