Underwater Image Enhancement Based on Histogram-Equalization Approximation Using Physics-Based Dichromatic Modeling

Yan-Tsung Peng, Yen-Rong Chen, Zihao Chen, Jung-Hua Wang, Shih-Chia Huang

TL;DR

This paper introduces a novel underwater image enhancement technique that combines the contrast-boosting power of Histogram Equalization (HE) with the color restoration capabilities of Physics-Based Dichromatic Modeling (PDM). By formulating this combination as a convex optimization problem, the method effectively addresses the common issues of low contrast and color distortion in underwater imagery, producing visually pleasing and more informative results.

What Problem Does This Paper Address?

Underwater environments present significant challenges for image capture. Light scatters and attenuates rapidly, leading to images that are often dark, hazy, and suffer from severe color casts (typically blue or green). These degraded images hinder the effectiveness of computer vision applications like object detection and semantic segmentation, which are crucial for underwater exploration and research. While existing methods can improve contrast or restore color, they often struggle to achieve both simultaneously, or they introduce unwanted artifacts. This paper aims to bridge this gap by developing a unified approach that enhances both contrast and color fidelity in underwater images.

Methods

The core of this work lies in a novel approach that synergizes two established image processing techniques:

  • Histogram Equalization (HE): A standard method for enhancing image contrast by redistributing pixel intensity levels to cover the entire dynamic range. However, HE alone can distort colors, especially in already color-cast underwater images.
  • Physics-Based Dichromatic Modeling (PDM): This approach models the image formation process in underwater environments, considering light scattering and absorption. PDM is effective at restoring natural colors but doesn't inherently guarantee good contrast.

The authors propose to combine these by formulating an objective function that balances the HE transformation with the PDM's color restoration. This is achieved through convex optimization. Specifically, they approximate the HE transformation function and then integrate it with a PDM-based term that models the distribution of restored pixel intensities. The optimization problem is designed to find a transformation that simultaneously enhances contrast (via the HE approximation) and preserves/restores color (via the PDM term), while also ensuring the transformation is physically plausible. The optimization is solved using the CVX package.

Datasets

The primary dataset used for evaluation is the UIEB (Underwater Image Enhancement Benchmark) dataset. This dataset comprises 890 real-world underwater images with corresponding high-quality reference images, covering various degradation types like haze, greenish, and blueish casts.

Key Insights and Ideas

  • Synergy of HE and PDM: The paper demonstrates that combining HE and PDM, rather than using them in isolation, is crucial for achieving both enhanced contrast and accurate color restoration in underwater images.
  • Convex Optimization as a Unifier: Framing the combined enhancement as a convex optimization problem provides a principled way to balance the competing goals of contrast enhancement and color fidelity.
  • Physics-Informed Enhancement: By incorporating PDM, the method leverages the underlying physics of underwater image formation, leading to more natural and robust results compared to purely data-driven or heuristic approaches.
  • Addressing Limitations of Existing Methods: The authors highlight that many existing methods either fail to restore color distortions or produce low-contrast images, and their proposed method directly tackles these shortcomings.

Key Results

  • Qualitative Improvement: The proposed method consistently produces visually pleasing underwater images with improved contrast and natural color balance, outperforming several state-of-the-art methods in visual quality. Examples show effective removal of haze and color casts without over-saturation.
  • Quantitative Performance: On the UIEB dataset, the proposed method achieved the best scores for the UCIQE (Underwater Color Image-Quality Evaluation) metric, indicating superior visual perceptual quality. It also achieved competitive scores for the NIQE (Natural Image Quality Evaluator) metric.
  • Feature Matching Application: An application to feature matching demonstrates that the enhanced images lead to a significantly higher number of matched feature points when compared to original or other enhanced images, suggesting improved image utility for downstream tasks.
  • Runtime: The proposed method has an average runtime of 4.8 seconds per image on the UIEB dataset, which is competitive, though not the fastest, compared to other methods.

Limitations and Future Work

  • Runtime: While competitive, the 4.8-second runtime might still be a bottleneck for real-time applications. Future work could focus on optimizing the algorithm for faster execution.
  • Generalization: Although the UIEB dataset is diverse, the method's performance on extremely varied or novel underwater conditions might require further investigation.
  • Parameter Tuning: The convex optimization involves a control parameter ($\alpha$) that needs to be appropriately tuned. While the paper suggests a range, finding the optimal value for all scenarios might require further study.

The authors suggest that their method can be further improved by exploring more advanced optimization techniques or by integrating it with deep learning architectures for potentially faster and more robust performance.

Summary updated

BibTeX
@inproceedings{peng2022underwater,
  author = {Yan-Tsung Peng and Yen-Rong Chen and Zihao Chen and Jung-Hua Wang and Shih-Chia Huang},
  title = {{Underwater Image Enhancement Based on Histogram-Equalization Approximation Using Physics-Based Dichromatic Modeling}},
  booktitle = {Sensors},
  year = {2022},
  url = {https://openreview.net/forum?id=omNMa3rMV1},
}
Figure