Enhancing Underwater Image via Adaptive Color and Contrast Enhancement, and Denoising

I. INTRODUCTION

• The method addresses underwater color cast, contrast loss, and noise, which existing methods may not comprehensively resolve and may aggravate through contrast overstretching.Reported shortcomings of some existing results include inadequate contrast, inaccurate color correction, limited brightness improvement, and noise amplification.Source: §I · pp. 1–2
• The proposed method is intended to improve color and contrast while suppressing noise.The introduction also reports effectiveness in color correction, contrast enhancement, and sharpness promotion.Source: §I · p. 2
• The framework separates image components to reduce noise effects and sharpen edges blurred by forward-scattering.It exploits noise concentration in the high-frequency component to avoid noise amplification in the low-frequency component.Source: §I · p. 2
• ACCE uses a Heaviside function to divide image channels into subsets and applies different enhancement strategies to limit over-enhancement and oversaturation.The paper presents ACCE as an adaptive variational framework for color and contrast enhancement.Source: §I · p. 2
• A numerical approximation combining gradient descent with a pyramid-based strategy is designed to reduce computational complexity.The strategy is used to optimize the framework.Source: §I · p. 2
• The authors state that the method requires neither prior knowledge nor an optical process and can be generalized to other degraded images.The introduction also reports experiments on underwater images with different challenging scenes.Source: §I · p. 2

II. RELATED WORK

• Restoration-based methods invert an underwater image formation model, but their reliance on scene assumptions limits performance across varied underwater environments.The review notes that these methods may also offer limited color-correction advantages.Source: §II · p. 2
• Enhancement-based methods directly modify pixel values, but traditional approaches may neglect other factors contributing to visual degradation.The review covers stretching, fusion, low-light enhancement, and variational approaches.Source: §II · p. 2
• Deep learning methods learn image mappings from training data, while collecting sufficient effective real-world and synthetic underwater images is difficult.The review also notes their dependence on data-driven learning.Source: §II · pp. 2–3
• Unlike direct contrast enhancement, the proposed method works on low-frequency components to isolate noise effects and adapt enhancement to local conditions.The authors report less information loss and improved contrast, color, and brightness; they also state that the method does not rely on a physical model.Source: §II · p. 3

III. PROPOSED METHOD

• In challenging underwater scenes, ACE may leave color casts uncorrected, although its local-global contrast enhancement remains useful.Figure 2 juxtaposes raw images with enhanced results from ACE and the proposed method.Source: §III · p. 3, Figure 2 · p. 3
• ACCE-D enhances the low-frequency component with ACCE, soft-thresholds noise in the DoG-derived high-frequency component, then linearly combines both components.A bilateral filter extracts low-frequency information; DoG separates high-frequency detail.Source: §III · pp. 3–4
• The framework uses a color-corrected guide Cr, with per-channel RGB mean and MSE setting thresholds for histogram stretching; λ is often set to 2.3.The enhanced HSI saturation and intensity channels are constrained to [0, 1].Source: §III · p. 4
• Minimizing the initial variational model can over-enhance saturation or brightness, create brighter or darker areas, and lose valuable information.Figure 4 compares raw and guided images with results from minimizing (2) and (5).Source: §III · p. 5, Figure 4 · p. 5
• ACCE adds Gaussian weights and a Heaviside function to apply different enhancement strategies across channel subsets and limit over-enhancement.Gaussian weights favor intermediate values near 0.5 over values near 0 or 1; adjusting σ changes their distribution and the strength of contrast enhancement.Source: §III · p. 5
• The authors solve ACCE with gradient descent and use a Gaussian-pyramid strategy to reduce the cost of computing its operators.The pyramid estimates global information at lower resolution, then updates it layer by layer toward adjacent regions.Source: §III · pp. 6–7

• The experiments examine the influence of ACCE parameters and the effect of the denoising strategy.Source: §IV · p. 7
• The study compares ACCE-D with advanced techniques through qualitative and quantitative experiments.Source: §IV · p. 7
• The experiments also extend ACCE-D to other degradation scenarios.The section describes the experimental setup as Matlab R2016b on Windows 10, with a 3.6 GHz CPU and 8 GB RAM.Source: §IV · p. 7

• With α=0.05, contrast is good but saturation is low; α=0.4 over-enhances color, while α=0.25 gives more perceptual color performance.The reported suitable range is α∈[0.15, 0.3].Source: A. Performance of proposed method · p. 7
• As β increases, contrast rises: β≤0.15 leaves residual haze, while larger β removes haze but can produce noisy results; β∈[0.15, 0.35] is suitable.The following experiments use β=0.3, and other tested α and β values are reported as acceptable.Source: A. Performance of proposed method · pp. 7–8
• Reducing σ suppresses over-enhanced regions, but very small σ also affects the framework globally; the authors report better results at σ=0.03.Source: A. Performance of proposed method · p. 8
• Across 60 underwater images captured under varied conditions, ACCE-D improves color dynamics and visibility, including in extremely deteriorated scenes.Fig. 7 presents the raw images and corresponding enhanced results.Source: A. Performance of proposed method · p. 8, Figure 7 · p. 8
• High-frequency denoising reduces noise amplification while preserving more edges and details than results without that strategy.The authors also report higher PSNR and SSIM for ACCE-D, calculated against the original images because underwater ground truth is unavailable.Source: A. Performance of proposed method · p. 8, Figure 8 · p. 8

• ACE improves contrast but corrects color inaccurately; RCP and IBLA can leave fog, UMSR introduces artifacts and unnatural gray, L2UWE gives little attention to color deviation, and GNCE has limited contrast improvement.Source: B. Qualitative comparison · p. 9
• ACCE-D is reported to have the best visual quality, recovering color and improving scene clarity without oversaturating pixels, so details are retained or highlighted.The low-light seaweed example is cited as an illustration; Fig. 9 compares results across challenging underwater scenes.Source: Figure 9 · p. 9, B. Qualitative comparison · p. 10
• Canny edge comparisons show more visible edges reproduced by ACCE-D than by the original images, while RCP, IBLA, and UMSR appear to leave fog.Fig. 10 presents the images and their corresponding edge maps.Source: B. Qualitative comparison · p. 10, Figure 10 · p. 10
• A broad hue distribution alone can mislead: RCP's apparent red-hue diversity reflects red-channel over-enhancement, while ACCE-D shows more chromaticity diversity than the other compared methods.The passage describes polar hue histograms and color-deviation detection as assessment tools; Fig. 11 shows the histogram comparisons.Source: Figure 11 · p. 10, B. Qualitative comparison · p. 11

C. Quantitative comparison

• The assessment uses UCIQE, UIQM, CPBD, and PCQI; UCIQE and UIQM are described as comprehensive underwater-image quality indicators.CPBD scores blur detection probability, while PCQI measures patch-based contrast variation.Source: C. Quantitative comparison · pp. 11–12
• ACCE-D outperforms other methods in most cases, and its average scores on 200 randomly selected benchmark images indicate greater robustness.The broader test reports average UCIQE, UIQM, CBPD, and PCQI scores.Source: C. Quantitative comparison · p. 12
• Across degraded-image types, ACCE-D's UCIQE scores exceed 0.6 and its UIQM scores exceed 1.2; it receives the highest evaluation for CBPD and PCQI.Source: C. Quantitative comparison · p. 12

D. Other applications

• For foggy images, ACCE-D achieves the best dehazing effect and more natural color, while other methods show color deviation, residual haze, or retained haze color.The comparison also notes that L2UWE increases visible edges and details but appears to retain the haze color.Source: D. Other applications · p. 12
• For sandstorm images with more serious yellowish color cast, ACCE-D produces more natural color and higher clarity than the compared methods.RCP and UMSR partly correct color bias but do not eliminate scattering; IBLA aggravates the color cast.Source: D. Other applications · pp. 12–13
• For low-light images, ACCE-D balances noise and illumination while slightly improving color, unlike comparison methods with artifacts, dark areas, or amplified noise.UMSR raises exposure and reveals details but also has over-enhanced areas and amplified noise.Source: D. Other applications · p. 13

V. CONCLUSION

• ACCE-D separates high- and low-frequency components using DoG and bilateral filters, then soft-thresholds the high-frequency component to reduce noise while preserving edges.Its low-frequency component is enhanced by a variational framework intended to improve visibility and reveal details without excessive enhancement.Source: §V · p. 13
• Experiments on real-world underwater images with different challenge scenes support ACCE-D's effectiveness in contrast enhancement, color correction, and noise suppression.The paper reports qualitative and quantitative results as validating superiority over other advanced methods.Source: §V · p. 13
• Application tests verify that ACCE-D can be extended to enhance other kinds of degraded scenarios.Source: §V · p. 13