Abstract
Underwater photography is inherently challenging due to the selective absorption and scattering of light, which results in low contrast, reduced brightness, and significant color distortion. While deep learning has advanced image restoration, most models rely solely on the RGB color space, where channels are highly correlated. We present UIEC^2-Net, a novel framework that integrates both RGB and HSV color spaces. By utilizing an RGB pixel-level block for local denoising and an HSV global-adjust block with a "neural curve layer," the network restores natural colors and contrast. An attention mechanism intelligently fuses these outputs, achieving state-of-the-art performance on benchmark datasets.
Introduction
As light travels through water, different wavelengths are absorbed at varying rates—red light disappears first, followed by orange and yellow—leaving images with a characteristic blue or green cast. Standard Convolutional Neural Networks (CNNs) typically operate in the RGB space. However, because RGB channels are coupled, adjusting one to fix a color cast often inadvertently alters the image's overall luminance or saturation. Properties like brightness (Value) and saturation are more effectively manipulated in the HSV space. UIEC^2-Net bridges this gap by implementing a differentiable pipeline that leverages the complementary strengths of both color representations.
Methodology
The UIEC^2-Net architecture consists of three primary modules designed to be trained end-to-end:
- RGB Pixel-Level Block: An 8-layer CNN focused on local pixel relationships to perform fundamental tasks such as denoising and initial color correction.
- HSV Global-Adjust Block: This module converts the RGB output into the HSV space. It employs a Neural Curve Layer—learned piecewise linear functions—to globally adjust saturation and value. This mimics professional "Curves" adjustments in photo editing but is automated through deep learning.
- Attention Map Block: To prevent artifacts caused by the high sensitivity of the Hue (H) channel, this block learns a pixel-level weight map. It selectively fuses the RGB and HSV outputs, ensuring the final image retains the best features of both while suppressing potential color noise.
To enable end-to-end training, the authors developed differentiable versions of RGB-to-HSV and HSV-to-RGB conversions using piecewise linear approximations, allowing gradients to flow through the entire system.
Experimental Results
The model was trained on a hybrid dataset comprising the UIEBD (800 real-world pairs) and synthetic data derived from the NYU-v2 RGB-D dataset (1,200 pairs). Testing was conducted across diverse water types to ensure generalization.
Quantitative Performance
UIEC^2-Net demonstrated superior performance on synthetic data, achieving a PSNR of 20.54 and an SSIM of 0.84, significantly outperforming previous models like UGAN and DUIENet. On real-world images, it consistently scored highest on no-reference metrics such as UCIQE and UIQM, which evaluate colorfulness, sharpness, and contrast.
Qualitative Analysis
Visual comparisons highlight the model's ability to eliminate the pervasive "background color cast" often left behind by other enhancement methods. While many models improve foreground clarity, UIEC^2-Net restores a natural, neutral tone to the entire scene, including distant backgrounds.
Conclusion
UIEC^2-Net proves that incorporating multiple color spaces can significantly enhance the effectiveness of underwater image restoration. By separating global adjustments (HSV) from local feature extraction (RGB), the model achieves a more balanced and natural restoration. Future work may involve replacing the current "plain" CNN backbone with more advanced architectures like U-Net or ResNet to further improve texture handling in complex underwater environments.
Summary updated