Abstract
Underwater imagery is increasingly drawing attention from the scientific community, since pictures and videos are invaluable tools in the study of the vast unknown oceanic environment that covers 90% of the planetary biosphere. However, underwater sensor networks must cope with the harsh channel that seawater constitutes. Medium range communication is only possible using acoustic modems that have limited transmission capabilities and peak bitrates of only a few dozens of kilobits per second. These reduced bitrates force heavy compression on videos, yielding much higher levels of distortion than in other video services. Furthermore, underwater video users are ocean researchers, and therefore their quality perception is also different from the generic viewers that typically take part in subjective quality assessment experiments. Computational efficiency is also important since the underwater nodes must run on batteries and their recovery is very expensive. In this paper, we propose a pixel-based no-reference video quality assessment method that addresses the described challenges and achieves good correlations against subjective scores of users of underwater videos.
Introduction: The Challenge of Underwater Video
Monitoring the 90% of our biosphere that lies underwater is a massive technological challenge. Underwater sensor networks (UWNs) rely on acoustic modems because electromagnetic waves (like Wi-Fi or LTE) are absorbed quickly by seawater. However, acoustic channels are incredibly narrow, offering peak bitrates of only a few dozen kilobits per second—thousands of times slower than a typical home internet connection.
To send video over these "thin" pipes, the data must be aggressively compressed, leading to severe artifacts. Standard Video Quality Assessment (VQA) tools fail in this environment for three primary reasons: the unique nature of artifacts from extreme compression, the specialized perception of ocean researchers who prioritize different visual information than casual viewers, and the severe resource constraints of battery-powered underwater nodes.
Proposed Methodology
The researchers developed a No-Reference (NR) VQA model, which estimates quality using only the distorted video without needing the original "pristine" version. The model is built on the theory of Natural Scene Statistics (NSS), which posits that natural, undistorted images follow predictable statistical distributions. When an image is distorted, it deviates from these statistics.
The model extracts six key features:
- Temporal Features ($f_1, f_2$): Calculates the difference between consecutive frames at two scales, measuring how much the distribution deviates from a natural Gaussian shape.
- Local Features ($f_3, f_4, f_5$): Breaks frame differences into patches to capture local distortions, grouping them into three levels based on statistical shape parameters.
- Spatial Feature ($f_6$): Utilizes the BRISQUE algorithm to assess the quality of individual frames, averaging the score over the entire video.
These features are processed by a Support Vector Regressor (SVR) with a Gaussian kernel to predict a final quality score.
Experimental Setup and Datasets
The model was trained and validated using a specialized underwater video database containing 31 different scenes, including marine fauna, flora, rocks, and the seafloor.
Videos were compressed using H.264 at bitrates between 8 and 20 kb/s and frame rates between 1 and 10 fps. Ground truth subjective scores (Mean Opinion Scores or MOS) were provided by 20 ocean scientists from the Spanish Oceanographic Institute, ensuring the model aligns with expert perception.
Key Results and Insights
The proposed model significantly outperformed existing state-of-the-art VQA algorithms:
- Proposed Model: Achieved a Linear Correlation Coefficient (LCC) of 0.81 and a Spearman Rank Order Correlation Coefficient (SROCC) of 0.76.
- V-BLIINDS: A popular general-purpose model reached only 0.50 LCC even after retraining.
- VIIDEO: A completely "blind" model failed almost entirely with an LCC of 0.12.
A key insight is that in low-bitrate underwater video, "jerkiness" and temporal artifacts are often more distracting than spatial blur. By focusing on frame differences, the model captures this temporal degradation effectively. Furthermore, the use of only six statistical features makes the model computationally efficient enough for energy-constrained underwater nodes.
Conclusion and Future Work
While the results are promising, the database size remains relatively small. Future work will focus on expanding the dataset and addressing specific underwater optical issues like backscatter or extreme color shifts. Additionally, further reducing feature extraction complexity could enable real-time execution on simple underwater microcontrollers for dynamic network resource management.
Summary updated