PaperlayerView exact revision on arXiv

Source-linked AI summary

Enhancement of Underwater Images with Statistical Model of Background Light and Optimization of Transmission Map

Wei Song, Yan Wang, Dongmei Huang, Antonio Liotta, Cristian Perra

arXiv:1906.08673v1eess.IVcs.MM

TL;DR

Underwater images suffer degradation from light absorption, scattering, and turbidity, making accurate physical restoration difficult. The paper combines statistical background-light estimation, optimized transmission maps, and white-balance color correction, and reports effective enhancement across varied underwater images while identifying a boundary for background-light estimation.

  • Problem

    Underwater restoration requires accurate background-light and transmission-map estimation because absorption and scattering severely degrade image quality.

  • Method

    The method builds a manually annotated background-light database, estimates background light statistically, optimizes channel-specific transmission maps, restores images physically, and applies white-balance correction.

  • Results

    Comprehensive evaluations report high background-light accuracy, rational low-complexity transmission estimation, and superior enhancement performance against state-of-the-art methods.

  • Takeaways & Limitations

    The method supports enhancement of various underwater images by combining image restoration with color correction tailored to underwater-image characteristics.

  • Takeaways & Limitations

    High-quality background lights were not obtainable for some close-shot objects because manual annotation was based on far-scene backgrounds.

Abstract

from arXiv · show

Underwater images often have severe quality degradation and distortion due to light absorption and scattering in the water medium. A hazed image formation model is widely used to restore the image quality. It depends on two optical parameters: the background light and the transmission map. Underwater images can also be enhanced by color and contrast correction from the perspective of image processing. In this paper, we propose an effective underwater image enhancement method for underwater images in composition of underwater image restoration and color correction. Firstly, a manually annotated background lights (MABLs) database is developed. With reference to the relationship between MABLs and the histogram distributions of various underwater images, robust statistical models of BLs estimation are provided. Next, the TM of R channel is roughly estimated based on the new underwater dark channel prior via the statistic of clear and high resolution underwater images, then a scene depth map based on the underwater light attenuation prior and an adjusted reversed saturation map are applied to compensate and modify the coarse TM of R channel. Next, TMs of G-B channels are estimated based on the difference of attenuation ratios between R channel and G-B channels. Finally, to improve the color and contrast of the restored image with a natural appearance, a variation of white balance is introduced as post-processing. In order to guide the priority of underwater image enhancement, sufficient evaluations are conducted to discuss the impacts of the key parameters including BL and TM, and the importance of the color correction. Comparisons with other state-of-the-art methods demonstrate that our proposed underwater image enhancement method can achieve higher accuracy of estimated BLs, less computation time, more superior performance, and more valuable information retention.

I. INTRODUCTION

Underwater enhancement is difficult because turbidity, absorption, scattering, and wavelength-selective attenuation degrade image quality, while existing restoration methods may be inaccurate or computationally complex. The paper combines statistical background-light estimation, optimized transmission maps, image restoration, and color correction, supported by a manually annotated database and comprehensive evaluation.

  • Motivation: Underwater image quality degradation arises from turbidity, light absorption, scattering, and wavelength-dependent attenuation, producing low-contrast and green-blue imagery.
  • Limitations of Existing Methods: Existing restoration methods can estimate background light and transmission maps inaccurately or require high computational complexity across diverse underwater scenes.White objects, foreground particles, ambient light, dim backgrounds, and selective attenuation can disrupt parameter estimation.
  • Contributions: The proposed method establishes a 500-image manually annotated background-lights database for underwater background-light estimation.The paper describes it as the first database for this task and reports accurate recovery using the annotations.
  • Contributions: A statistical background-light model combines channel-wise means, medians, and standard deviations without requiring prior image information.The authors report improved accuracy and substantially reduced computation time compared with recent models.
  • Contributions: Transmission maps use a new underwater dark channel prior, depth-map compensation, adjusted reversed saturation optimization, and channel-specific attenuation differences.The red-channel map is optimized first, while green-blue maps account for their different attenuation rates.
  • Contributions: The enhancement pipeline restores images with the image formation model and then applies modified white balance to improve color, contrast, and natural appearance.The paper evaluates background light, transmission maps, color correction, and comparisons with state-of-the-art methods.

II. RELATED WORK

Related restoration methods estimate underwater background light and transmission through physical image-formation models, dark-channel variants, channel differences, and blurriness-based candidate selection. These approaches vary in assumptions, candidate construction, and computational procedures.

  • Physical Image Formation: The simplified image formation model expresses hazed intensity as direct radiance transmission plus background-light scattering.It uses separate red, green, and blue channel parameters.
  • Physical Image Formation: Transmission depends exponentially on attenuation and scene depth, with normalized residual energy ratios varying by wavelength and water type.For ocean type I water, the cited residual-energy ranges are 80%–85% for red, 93%–97% for green, and 95%–99% for blue.
  • Background-Light Estimation: Traditional background-light methods select brightest pixels, dark-channel extrema, channel differences, or regions scored by average intensity and variance.These methods can be affected by bright foreground objects, suspended particles, and channel-specific attenuation.
  • Background-Light Estimation: Quad-tree methods recursively search rectangular regions for flat backgrounds using scores based on average intensity and standard deviation.A final background light is then estimated from the selected region or candidate values.
  • Blurriness-Based Estimation: Multiple-candidate methods derive candidates from blurry pixels, low-variance regions, and highly blurry regions using an initial blurriness map.Gaussian filtering, maximum filtering, morphological reconstruction, guided filtering, and region subdivision are used in the described procedure.

C. TM Estimation

Transmission-map estimation builds on dark-channel priors while adapting to underwater attenuation and scene-depth characteristics. The described approach combines channel-specific priors, depth-related corrections, and refinement operations to obtain more suitable transmission maps.

  • Dark-Channel Priors: The dark channel prior assumes that most non-sky local patches contain at least one RGB pixel with extremely low intensity.
  • Transmission Derivation: Transmission estimation applies minimum filtering to the image-formation model and substitutes the dark-channel prior to derive a transmission map.
  • Dark-Channel Priors: Underwater variants restrict the dark channel to green-blue channels because red light is aggressively attenuated.The underwater dark channel prior replaces RGB-channel intensity terms with green-blue terms where specified.
  • Transmission Derivation: The maximum intensity prior estimates transmission from the difference between maximum red intensity and maximum green-blue intensity.
  • Depth-Based Refinement: Scene depth can combine stretched red-channel maximum, maximum-intensity, and image-blurriness maps before calculating the red-channel transmission.The described formulation refines relative depth with a guided filter and transforms it into actual distance.
  • Channel-Specific Estimation: Green-blue transmission maps are obtained using their attenuation ratios relative to the red channel.

III. OUR PROPOSED METHOD

The proposed method combines statistical background-light estimation, transmission-map optimization, physics-based restoration, and white-balance color correction for diverse underwater images.

  • Overall method: The method estimates RGB background lights statistically, then estimates transmission maps and applies them in an underwater image formation model.Its flow combines a manually annotated background-lights database with transmission-map estimation and optimization.
  • Color correction: The restored image receives final brightness and color correction through modified white balancing with an optimal gain factor.This post-processing step is intended to improve contrast, color, and natural appearance after restoration.
  • MABLs database: The MABLs database contains manually annotated background-light values for 500 underwater images selected from diverse scenes and distortions.More than 2,000 images were collected, 1,000 were screened for diversity, and unreliable annotations were removed.
  • MABLs database: Background-light annotations select far, high-intensity scene points and retain the point with the largest R-versus-GB intensity difference.This selection follows the stronger attenuation of red light compared with green and blue light underwater.
  • BL estimation model: The statistical model links MABLs to RGB histogram characteristics and uses separate linear and nonlinear models for channel-specific background-light estimation.The paper reports that existing methods can be computationally expensive and scene-specific, motivating this model.

New UDCP

The proposed background-light model uses relationships between underwater histograms and manually annotated lights, with regression analysis selecting channel-appropriate statistical relationships.

  • Histogram relationships: Histograms from five typical underwater scenes illustrate how RGB distribution trends relate to manually annotated background lights.The histogram lines represent the probability distributions of the red, green, and blue channels.
  • Histogram relationships: Histogram medians can closely match annotated background lights and avoid some white-region errors affecting DCP- and UDCP-based estimation.Other examples show green, red, or near-balanced channel dominance in the background light.
  • Regression analysis: The paper uses Pearson correlation for G-B channels and Spearman correlation for the nonnormally distributed R-channel relationships.The choice reflects linear relationships in G-B channels and monotonic but non-linear relationships in the R channel.
  • Regression analysis: The estimated background lights are clipped to the interval 5 to 250 to limit over-fitting or under-fitting from the limited MABLs.The clipping is implemented as B_c = min(max(B_c′,r, 5), 250) for c ∈ {r, g, b}.
  • Regression analysis: Adjusted R2 exceeds 0.6 for the linear model, while R2 exceeds 0.65 for the nonlinear model.The paper states that these models account for about 60% of the variability in background lights.

B. TM Estimation Model

The paper estimates underwater transmission maps using a statistically derived dark-channel prior, then compensates and optimizes the red-channel map with depth and saturation cues before estimating green-blue maps.

  • Red-channel transmission map: DCP and UDCP assumptions can produce erroneous transmission maps because they ignore wavelength-dependent underwater attenuation.Errors arise for scenes containing white fish, dim red-channel regions, or artificial foreground lighting.
  • New underwater dark channel prior: The prior is derived from 1000-image R-G-B and 2000-image G-B dark-channel histograms and their cumulative distributions.Each histogram bin represents 16 intensity levels over the 0–255 range.
  • New underwater dark channel prior: The new underwater dark-channel prior is set from high-quality image statistics rather than assuming a zero dark-channel value.The R-G-B dark-channel distribution contains approximately 40% zero values, while the proposed prior uses 25 after normalization.
  • Red-channel transmission map: The coarse red-channel transmission map is compensated with an underwater light attenuation depth map and optimized using an adjusted reversed saturation map.The depth cue uses the difference between maximum G-B intensity and R intensity, while artificial-light regions are associated with low HSV saturation.
  • Transmission-map refinement: After compensation and optimization, inaccurate red-channel transmission maps are rectified, then refined with guided filtering and histogram stretching.The recovered image applies the refined transmission map with RGB background-light estimates, using transmission bounds of 0.2 and 0.9.

C. Color Correction

The proposed color correction applies a white-balance algorithm with an optimal gain factor to improve brightness, contrast, and natural color appearance after restoration. Experiments select λ_v=0.25 as the preferred setting.

  • Color correction: The color correction applies modified white balance with an optimal gain factor to the coarse underwater image after restoration.It is intended to address low brightness and contrast that can obscure image details.
  • Parameter selection: λ_v is selected between 0 and 0.5, with values closer to 0 producing lower corrected-image brightness.The parameter controls the desired color and brightness of the enhanced image.
  • Parameter selection: λ_v=0.25 is chosen as optimal because λ_v=0.2 appears slightly bright whereas λ_v=0.3 appears somewhat dim.The selected setting preserves the actual color tone while improving visibility and contrast.
  • Parameter selection: Figure 11 compares the original image with enhanced images produced using λ_v values from 0 to 0.5.The displayed settings are 0, 0.1, 0.2, 0.25, 0.3, 0.4, and 0.5.

IV. RESULTS AND EVALUATION

For fair evaluation, all test images are resized to 400×600 pixels, and implementations run under a specified Windows 7 and Python 3.6.3 environment.

  • Experimental setup: All test underwater images are pre-processed to 400×600 pixels for evaluation fairness.The experiments run on a Windows 7 PC with an Intel Core i7-4790U CPU, 8 GB memory, and Python 3.6.3.

A. Evaluation of Objectives and Approaches

The evaluation examines the proposed MABLs, statistical background-light model, transmission-map optimizer, and overall enhancement method through comparisons and multiple quality metrics.

  • Evaluation objectives: The experiments assess MABLs, the statistical BL model, the TM optimizer, and the overall underwater image enhancement method.These four objectives cover both individual components and the complete approach.
  • Compared methods: Comparisons include MIP, DCP, underwater optical-model, UDCP, blurriness-based depth, Red channel, Li’s, and Peng’s methods.The evaluation compares underwater restoration and enhancement approaches.
  • Evaluation metrics: The evaluation uses RMSE, SSIM, Entropy, BRISQUE, UCIQE, and running time to assess restoration and enhancement performance.RMSE and SSIM are full-reference metrics, while the other listed image-quality metrics are used as non-reference measures.

B. Effectiveness of Proposed MABLs

The MABLs database contains 500 underwater images annotated using far scene points with high intensity and strong R-versus-G-B differences. Its effectiveness is evaluated against several estimated BLs under common transmission maps, with correctness limited for close shots.

  • Effectiveness of Proposed MABLs: The MABLs database contains 500 underwater images annotated from far scene points with high intensity and large R-channel versus G-B-channel differences.The database is used to examine whether the manually selected background lights are appropriate.
  • Effectiveness of Proposed MABLs: MABLs are compared with BLs from DCP, MIP, UDCP, Li’s, Peng’s, and the proposed method while restoring images with the same TMs.Restored-image quality is then evaluated under the different background-light choices.
  • Limitations: The MABLs cannot be guaranteed absolutely correct for close-shot images where the far scene point is difficult to identify.For these images, MABLs may not significantly improve quality but at least do not cause distortion.
  • Quantitative analysis: Table II reports RMSE, SSIM, Entropy, BRISQUE, UCIQE, and running time for restoration results using different BLs and all TMs.The table provides quantitative analysis of the BL comparison.
  • Qualitative results: Figure 12 presents successful results based on the MABLs with the TMs from Peng’s method.

C. Statistical Model of BL Estimation

The statistical BL model is evaluated against manually annotated references for accuracy and efficiency, alongside visual comparisons using different BL estimators. It achieves higher prediction accuracy and substantially lower running time than the compared methods.

  • Evaluation setup: BL predictions are evaluated against MABLs using 30% testing data, with accuracy tolerances of 30 for R and 40 for G-B channels.The evaluation covers 500 images and labels predictions as accurate when they fall within the specified channel tolerances.
  • Quantitative performance: The proposed statistical model outperforms five comparison methods in image-quality assessment and has significantly lower running time.The authors report that BL estimation can save 1000% of time relative to the fastest UDCP method.
  • Quantitative performance: DCP and UDCP produce the two lowest BL-estimation accuracies among the evaluated methods.The other three comparison methods also have visibly lower overall prediction accuracy than the proposed statistical model.
  • Efficiency: DCP, UDCP, Peng, and Li exhibit running times that increase linearly with image size, while MIP is slowest because maximum intensity is computed twice.The proposed non-linear R-channel and linear G-B-channel models require limited time to determine BLs.
  • Visual comparison: Using different BLs with the same TM produces visibly different restored color tones across the compared methods.The comparison includes DCP, MIP, UDCP, Li, Peng, the proposed method, and MABLs.

D. Performance of TM Optimizer

The TM optimizer is assessed through restored-image comparisons and quantitative quality measures using MABLs. Channel-specific transmission maps improve information, contrast, and local detail, while the proposed method achieves performance comparable to Peng’s with nearly half its running time and benefits from color correction.

  • TM design: DCP, MIP, and UDCP use one TM for all channels, ignoring the difference between the R-channel and G-B-channel TMs.This produces over-saturated areas in restored coral and fish images.
  • TM design: Separate R and G-B TMs increase valuable information, overall contrast, and local details in the restored images.The result is reported for Peng’s method and the proposed method, unlike the over-saturation observed with DCP, MIP, and UDCP.
  • Comparison design: Restored results based on different TMs are compared using MABLs, while the overall method is compared with multiple restoration and enhancement baselines.The TM comparison includes DCP, MIP, UDCP, Peng, and the proposed method; the broader comparison also includes Li, OWCC, and histogram-equalized variants.
  • Quantitative comparison: Peng’s method is slightly better in SSIM and BRISQUE, while its running time is nearly double that of the proposed method.The full-reference and non-reference analyses compare RMSE, SSIM, Entropy, BRISQUE, UCIQE, and running time.
  • Color correction: The method combines restoration with white-balance color correction to improve contrast, preserve genuine color, and improve visibility.The correction is applied after restoration using an optimal gain factor, and the evaluation includes methods with and without post-processing.
  • Color correction: Adding histogram equalization improves quality assessment for DCP, MIP, UDCP, and Peng, but slightly increases running time.The overall restoration and enhancement comparison reports RMSE, SSIM, Entropy, BRISQUE, UCIQE, and RT.
  • Efficiency: For an image with x∙y pixels, the reported running time is approximately 0.71∙(x∙y)/(400∙600) seconds.The estimate is based on images with 400×600 pixels processed by the proposed method.
  • Color correction: Color correction reduces greenish illumination and improves brightness, color, and contrast in the comparison results.The final proposed images are described as neither over-saturated nor over-enhanced, with objects better differentiated from the background.

V. DISCUSSION

The discussion presents the method across challenging underwater scenes and compares it with CNN-based restoration under different experimental conditions. It reports robust BL estimation and improved enhanced images, while acknowledging limitations for close-shot objects and manually annotated BLs.

  • V. DISCUSSION: The method is tested on greenish, bluish, thickly hazed, low-visibility, turbid, and dim underwater scenes.These challenging cases are presented as examples of the method’s restoration performance.
  • V. DISCUSSION: White balancing with an optimal gain factor is essential for improving the color and contrast of dehazed images because restoration alone removes haze and blur.The discussion distinguishes restoration’s dehazing role from post-processing’s color and contrast improvement.
  • V. DISCUSSION: The statistical BL model estimates global scene background light while avoiding interference from white blocks or bright foreground points.This behavior is reported for several challenging scenes in Fig. 17.
  • V. DISCUSSION: The CNN comparison uses different experimental conditions, and the authors note that CNN training requires a huge amount of data.The comparison uses extracted underwater images with low original resolution.
  • VI. CONCLUSION: The method uses a MABLs database, channel-specific BL statistics, an NUDCP-based R-channel TM, depth and saturation-map adjustments, and optical estimation of G-B TMs.Improved white-balance color correction is then used to enhance restored-image contrast and color.
  • VI. CONCLUSION: Comprehensive evaluations report effective MABLs, accurate BL estimation, rational and low-complexity TM estimation, and superior overall enhancement performance.The comparison covers other state-of-the-art restoration and enhancement methods.
  • VI. CONCLUSION: The method could not obtain high-quality BLs for some close-shot objects because of manual BL annotation based on far scenes.For these images, the authors report that restoration is not sensitive to the BLs and does not further distort the original image.
Loading 1906.08673v1…