Technical Summary: Beyond Aesthetics: Quantifying Information Loss in Turbid Scenes

Problem Statement

Underwater computer vision systems face significant performance degradation in turbid environments due to light attenuation and scattering. While these perceptual degradations (color distortion, contrast loss, blurring) are well-documented, the specific impact on downstream machine learning models remains unclear. A critical gap exists in the field: there are no thoroughly annotated real-world datasets covering the full spectrum of turbidity, forcing researchers to rely on synthetic data. However, existing synthetic datasets often focus on perceptual enhancement (e.g., color correction) rather than the structural information loss that actually hinders model performance. Furthermore, standard image quality metrics (e.g., PSNR, SSIM, NIQE) often fail to correlate with model performance in these conditions, as they prioritize human perception over the structural information required for machine learning tasks.

Methodology

1. The Turbid Underwater Baseline (TUB) Dataset

To address the lack of real-world data, the authors introduced the TUB dataset.

• Collection Protocol: Data was captured in a controlled cylindrical tank using four static GoPro cameras. Objects (natural and man-made) were fixed to a baseplate to ensure scene consistency across varying turbidity levels.
• Turbidity Control: Turbidity was induced using measured amounts of oat milk, creating 10 distinct levels per scene. A nephelometer was used to quantify turbidity in Nephelometric Turbidity Units (NTU), categorizing images into Low (0–10 NTU), Medium (10–20 NTU), and High (20+ NTU) groups.
• Annotations: The dataset comprises 1,320 images across 33 unique scenes. Crucially, instance segmentation masks were manually annotated on the clear-water images and propagated to all corresponding turbid images. This approach yields over 16,000 high-confidence ground-truth masks, even for scenes where human annotators could not visually distinguish object boundaries.

2. Proposed Metric: Phase-Congruency Delentropy (PCD)

The authors propose PCD, a metric designed to quantify structural information loss rather than perceptual quality.

• Theoretical Basis: The metric is built on the hypothesis that deep learning models rely on structural information (gradients and boundaries) which may persist even when contrast is lost.
• Mechanism: PCD calculates the delentropy (a measure of information complexity based on gradient distribution) but replaces standard kernel-derived gradients with phase congruency maps.
• Advantages: Phase congruency is invariant to illumination and contrast changes. By computing delentropy on these maps, PCD captures the loss of structural information (e.g., blurred edges) while remaining robust to the contrast reduction and color shifts typical of turbid water.

3. Experimental Evaluation

The study evaluated the correlation between various metrics and the performance of three state-of-the-art instance segmentation models: MaskRCNN, YOLOv11, and Mask2Former.

• Real Data: Models were trained and tested on splits of the TUB dataset (Low, Medium, High turbidity).
• Synthetic Data: The authors generated synthetic turbid images using two methods: a standard underwater image formation model (Synth1) and a model incorporating blur and inhomogeneity (Synth2).
• Comparison: The correlation of PCD against model performance (AP50) was compared against common metrics: PSNR, SSIM, NIQE, Entropy, Delentropy, UCIQE, and UIQM.

Key Results

1. Dataset Utility: The TUB dataset successfully provides the first high-confidence ground-truth annotations for extreme turbidity, enabling data-driven analysis of degradation effects.
2. Metric Correlation:
   • PCD demonstrated a strong correlation with the performance of instance segmentation models across both real and synthetic turbid images.
   • Existing Metrics: Common perceptual metrics (PSNR, SSIM, NIQE, UCIQE) showed weak to no correlation with model performance. Notably, these metrics often rated synthetic images as having lower quality than real turbid images, despite models performing well on the synthetic data.
3. Structural vs. Perceptual: The results indicate that a reduction in perceptual quality (e.g., low contrast) does not necessarily equate to a loss of task utility. Conversely, the loss of structural information (gradient degradation) is the primary factor driving performance drops in detection models.
4. Synthetic Data Limitations: Standard synthetic pipelines failed to reproduce the specific structural degradations of real turbidity, leading to misleading evaluations when using traditional metrics.

Significance and Claims

The paper claims that the current reliance on perceptual metrics and synthetic data is insufficient for advancing underwater computer vision. The authors argue that:

• PCD offers a superior, model-centric approach to evaluating turbid scenes by focusing on the structural information that learning-based models actually utilize.
• The TUB dataset fills a critical gap by providing a benchmark for extreme turbidity with reliable ground truth, allowing for the first direct assessment of how turbidity affects learning-based models.
• The findings suggest that improving "perceptual quality" (making images look clearer to humans) does not guarantee better model performance; instead, preserving or recovering structural gradients is paramount.

The work concludes that PCD and TUB provide essential tools for quantifying information loss and evaluating the realism of synthetic turbidity generation, moving the field beyond aesthetic improvements toward functional robustness in underwater environments.