Source-linked AI summary
Underwater Single Image Color Restoration Using Haze-Lines and a New Quantitative Dataset
Dana Berman, Deborah Levy, Shai Avidan, Tali Treibitz
TL;DR
Underwater restoration must address color- and distance-dependent degradation that varies with water properties, while existing evaluations provide limited quantitative evidence. The paper uses a wavelength-aware single-image method that tests water-type attenuation ratios and introduces an in situ stereo dataset with color-chart ground truth. The authors report quantitative comparisons showing their algorithm is competitive and that it outperforms other algorithms in transmission accuracy.
Problem
Underwater images suffer wavelength- and distance-dependent color distortion and contrast loss, while existing methods often neglect diverse spectral properties and lack comprehensive color-aware evaluation.
Method
The method reduces wavelength-dependent restoration to single-image dehazing by estimating two global attenuation ratios across candidate Jerlov water types and selecting the best color distribution.
Results
The algorithm is competitive with other state-of-the-art methods, and outperforms other algorithms in transmission accuracy.
Takeaways & Limitations
The wavelength-aware model reconstructs complex 3D underwater scenes and corrects colors of distant objects, while the stereo dataset enables quantitative evaluation on natural images.
Takeaways & Limitations
Veiling-light estimation assumes a visible object-free smooth area, an assumption that fails for close reef walls or downward-facing cameras.
Abstract
from arXiv · showhide
Underwater images suffer from color distortion and low contrast, because light is attenuated while it propagates through water. Attenuation under water varies with wavelength, unlike terrestrial images where attenuation is assumed to be spectrally uniform. The attenuation depends both on the water body and the 3D structure of the scene, making color restoration difficult. Unlike existing single underwater image enhancement techniques, our method takes into account multiple spectral profiles of different water types. By estimating just two additional global parameters: the attenuation ratios of the blue-red and blue-green color channels, the problem is reduced to single image dehazing, where all color channels have the same attenuation coefficients. Since the water type is unknown, we evaluate different parameters out of an existing library of water types. Each type leads to a different restored image and the best result is automatically chosen based on color distribution. We collected a dataset of images taken in different locations with varying water properties, showing color charts in the scenes. Moreover, to obtain ground truth, the 3D structure of the scene was calculated based on stereo imaging. This dataset enables a quantitative evaluation of restoration algorithms on natural images and shows the advantage of our method.
1 INTRODUCTION
Underwater light attenuation and scattering produce distance-dependent color distortion and contrast loss, while existing single-image methods often overlook diverse water spectra and lack quantitative color evaluation. The paper proposes single-image recovery that models optical water types and evaluates restoration against stereo-derived ground truth.
- Underwater scattering and absorption distort colors and reduce contrast, limiting visual surveys and downstream tasks such as segmentation, feature matching, and navigation.The degradation varies with wavelength and object distance, so it cannot be globally corrected.
- Existing single-image enhancement methods underperform because they ignore diverse water spectral properties and are usually evaluated qualitatively on few images.No-reference metrics used by some methods focus on luminance and cannot measure color correction.
- The proposed method recovers distance maps and object colors from one underwater image using a more comprehensive physical image-formation model across optical water types.Single-image recovery avoids requiring additional equipment while accommodating temporally changing water properties.
- Color-dependent transmission reduces to four per-pixel unknowns by estimating two global attenuation-ratio parameters, selected from a library of water types using corrected-image color distributions.The reduction exploits the relation between channel transmission and distance.
- The authors collected stereo images containing color charts, recovered true camera distances, and quantitatively compared multiple algorithms against ground truth.Their algorithm is reported as competitive with other state-of-the-art methods.
2 RELATED WORK
Prior underwater methods use multiple images, active equipment, simplified physical assumptions, or visually oriented enhancement, while evaluations often lack realistic 3D ground truth. The paper addresses these gaps with wavelength-aware modeling and a quantitative in situ dataset.
- Multi-image and active-illumination approaches can estimate attenuation or improve visibility, but their equipment, viewpoint, or distance requirements limit applicability.
- Single-image dehazing methods commonly assume color-independent transmission, an assumption violated underwater because attenuation varies by wavelength.
- Some prior methods classify water globally or disregard spectral dependence, limiting compensation for distance-dependent attenuation.
- DCP-based underwater methods can misestimate transmission for bright foreground sand and dominant-color background water.
- Visually pleasing enhancement methods have not demonstrated the color consistency required for scientific measurements.
- Synthetic training data and tank experiments remain limited by inaccurate spectral simulation or restricted realism, motivating in situ evaluation.
3 BACKGROUND
The image-formation model represents underwater observations as attenuated scene radiance plus veiling light, with transmission governed by distance and wavelength-dependent attenuation. Jerlov water types constrain the attenuation ratios used to reduce restoration to single-image dehazing.
- 3.1 Image Formation Model: Each channel combines attenuated object radiance with a global veiling-light component, which represents scattered ambient light in object-free regions.
- 3.1 Image Formation Model: Transmission decreases exponentially with distance according to the channel-specific attenuation coefficient βc.
- 3.1 Image Formation Model: Red attenuation can be an order of magnitude larger than blue and green, making underwater transmission wavelength-dependent.
- 3.2 Water Attenuation: Jerlov classifies open-ocean and coastal waters by clarity, with types ranging from clearest to most turbid within each group.
- 3.2 Water Attenuation: The method constrains RGB attenuation using Jerlov types and estimates two ratios rather than the three absolute attenuation coefficients.
4 COLOR RESTORATION
The method evaluates candidate water types through their attenuation ratios, estimates transmission and veiling light from one image, and selects the restoration most consistent with Gray-World colors.
- Water-type compensation: Ten Jerlov water types provide candidate attenuation-ratio pairs for reducing color-dependent transmission to one common per-pixel transmission.The two global ratios are βBR = βB/βR and βBG = βB/βG.
- Veiling-light estimation: Veiling light is estimated from smooth, object-free image regions, whose average color supplies the ambient-light estimate A.Structured-edge detection identifies the largest connected component used as veiling-light pixels.
- Transmission estimation: Haze-Lines clustering estimates an initial blue-channel transmission, then a 0.9 factor accounts for the absence of truly haze-free underwater pixels.Jerlov measurements indicate that one-meter scene points can still have blue transmission around 0.9, depending on water type.
- Transmission estimation: A lower transmission bound follows from nonnegative restored radiance, while soft matting smooths transitions between veiling-light and object regions.Mahalanobis distance from the veiling-light distribution identifies candidate non-object pixels; soft matting avoids abrupt, nonphysical discontinuities.
- Color restoration: The restored image compensates path attenuation, applies global white balance to remove colored illumination, and converts linear data to sRGB.White balance addresses ambient illumination attenuated through the water column, while the final conversion uses color-space conversion and gamma tone mapping.
- Water-type selection: The algorithm repeats restoration across water types and selects the result with the smallest red, green, and blue mean difference under Gray-World; this measure outperformed maximal contrast.Gray-World is applied outside veiling-light pixels because open water without objects does not satisfy the assumption.
5 EXPERIMENTS
The experiments introduce an in situ stereo dataset with color charts and ground-truth distances, then compare underwater enhancement methods qualitatively and quantitatively. Results assess transmission accuracy, color restoration across distances, and sensitivity to the assumed water type.
- Experimental Set-up and Method: Existing evaluations often use short-range color cards, tanks, pools, or heuristic no-reference metrics, limiting assessment of natural 3D scenes and color correction.The proposed evaluation instead targets scenes with objects at different distances and varying natural-water properties.
- Experimental Set-up and Method: The dataset contains in situ scenes from different seasons, depths, and water types, with identical waterproof color charts placed at multiple camera distances.This setup tests whether restoration remains consistent as distance-dependent degradation changes.
- Experimental Set-up and Method: Stereo calibration, dense matching, and triangulation recover scene geometry; the stereo pair is used only to generate ground-truth distance maps, not for restoration.Color charts are masked when images are supplied to the single-image algorithm.
- Qualitative comparison: Qualitative comparisons show that some transmission-based methods inconsistently correct colors, while texture enhancement can create false foreground and sand textures.A color-transfer method is relatively effective across different distances, particularly on the color charts.
- Transmission Estimation - Quantitative Evaluation: The proposed method estimates more accurate transmission than competing algorithms, evaluated by correlating −log(t) with stereo-derived distances.The Pearson coefficient ranges from −1 to +1; negative values can indicate incorrect veiling-light estimation or an invalid prior.
- Color Restoration - Quantitative Evaluation: Color-restoration evaluation uses average reproduction angular error in RGB space, where lower angles indicate greater accuracy; the leading methods are the proposed method and.Global contrast adjustment helps chiefly for the nearest chart, while farther charts require distance-dependent correction.
- Color Restoration - Qualitative comparison: Using incorrect attenuation ratios for the water type produces distorted colors, whereas the automatically selected type C5 restores the murky-water image differently from open-ocean type I.This comparison demonstrates the practical importance of estimating water type.
6 CONCLUSIONS
The paper extends haze-lines restoration to wavelength-dependent underwater attenuation and introduces a real-world stereo dataset with color-chart ground truth for quantitative evaluation.
- 6 CONCLUSIONS: The method models multiple water types and adds only two global parameters for separate color-channel transmission recovery.This extends the haze-lines model to wavelength-dependent attenuation while preserving a compact estimation problem.
- 6 CONCLUSIONS: The physical model reconstructs scenes with complex 3D structure and corrects colors of distant objects.The conclusion specifically links the model to recovery in scenes whose depth structure and object distances vary.
- 6 CONCLUSIONS: The authors collected an in situ stereo dataset of underwater images with color charts and ground truth.The dataset is intended to support quantitative evaluation on natural underwater scenes and has implications for oceanic research.