Underwater Organism Color Enhancement via Color Code Decomposition, Adaptation and Interpolation

TL;DR

The paper tackles underwater color distortion by introducing ColorCode, a framework that decomposes an underwater image into a color code $m$ and a content code $c$ to enable deterministic color enhancement, guidance-based color adaptation, and continuous color interpolation. A supervised training stage recovers a reference image, while self- and cross-reconstruction disentangle color semantics from content, with explicit Gaussian constraints on $m$ and a color-guided fusion mechanism for long-wavelength hues. ColorCode delivers three capabilities—color enhancement, color adaptation, and color interpolation—validated across multiple benchmarks with quantitative gains and compelling visual results, outperforming several baselines in diversity, realism, and controllability. This approach enables flexible, user-controllable underwater image enhancement suitable for publication-ready visuals and downstream analysis.

Abstract

Underwater images often suffer from quality degradation due to absorption and scattering effects. Most existing underwater image enhancement algorithms produce a single, fixed-color image, limiting user flexibility and application. To address this limitation, we propose a method called \textit{ColorCode}, which enhances underwater images while offering a range of controllable color outputs. Our approach involves recovering an underwater image to a reference enhanced image through supervised training and decomposing it into color and content codes via self-reconstruction and cross-reconstruction. The color code is explicitly constrained to follow a Gaussian distribution, allowing for efficient sampling and interpolation during inference. ColorCode offers three key features: 1) color enhancement, producing an enhanced image with a fixed color; 2) color adaptation, enabling controllable adjustments of long-wavelength color components using guidance images; and 3) color interpolation, allowing for the smooth generation of multiple colors through continuous sampling of the color code. Quantitative and visual evaluations on popular and challenging benchmark datasets demonstrate the superiority of ColorCode over existing methods in providing diverse, controllable, and color-realistic enhancement results. The source code is available at https://github.com/Xiaofeng-life/ColorCode.

Figures (27)

• Figure 1: As the water depth increases from $d_{1}$ to $d_{4}$, light of various wavelengths gradually diminishes, causing underwater organisms to shift toward blue-green or blue hues, which appear dull, as shown in (a). To address this, we propose ColorCode, which offers three key features: (b) color enhancement $\mathcal{F}_{ce}(\cdot)$, (c) color adaptation $\mathcal{F}_{ca}(\cdot, \cdot, \cdot)$ using natural guidance $g_{1}$ or underwater guidance $g_{2}$, and (d) color interpolation $\mathcal{F}_{ci}(\cdot, \cdot)$ through sampling different color codes.
• Figure 2: Comparisons of results obtained by style code and color code. The color blocks represent the main colors of the organisms in the images. The acquisition of color block is achieved by calculating the center of mass of the organism and the surrounding main color.
• Figure 3: Observation of approximately invariance of hue of long-wavelength colors during enhancement. $x_1$ and $x_2$ represent two underwater images with different distortion levels, where $y_1$ and $y_2$ mean the corresponding reference images. $p-$ denotes the enlarged patch. The curve figure stands for the pixel histogram, while the bar figure shows the average pixel value of the red (R), green (G) and blue (B) channel. The images in (a) and (b) are from islam2020fast. In (a), there is no long-wavelength color. In (b), the long-wavelength color (yellow) is preserved during the enhancement from $x_{2}$ to $y_{2}$.
• Figure 4: The distribution histogram of each dimensions in color code, where the $x$ and $y$ axis coordinates represent the value range and amount, respectively. The red line represents the location of the mean of the data, while the green line denotes $x=0$.
• Figure 5: The overall training and inference process of the proposed ColorCode. The $P1$, $P2$, $P3$, and $P4$ stand for the enhancement (Section \ref{['subsec:enhancement_process']}), decomposition (Section \ref{['subsec:extraction_of_the_content_code']} and Section \ref{['subsec:explicit_constraint_on_the_color_code']}), color adaptation (Section \ref{['subsec:color_guidance_by_the_image_with_long_wavelength_color']}), and color code interpolation (Section \ref{['subsec:color_interpolation_by_sampling_color_codes']}) processes, respectively. For the meanings of other symbols, please refer to Section \ref{['sec:methods']}.