TLDR: This paper introduces an improved synthetic underwater image generation model that addresses the limitations of traditional models, especially in highly turbid environments. By reintroducing the forward scattering term and incorporating medium inhomogeneity using Gaussian Random Fields, the proposed pipeline generates significantly more realistic underwater images. The researchers also created the BUCKET dataset under controlled turbidity conditions to provide ground truth for evaluation. A survey confirmed the qualitative improvements, with participants preferring the new model’s outputs for realism, particularly in turbid coastal waters.
Underwater computer vision is becoming increasingly important for monitoring our marine ecosystems, which are under threat from climate change and pollution. However, collecting data underwater is incredibly challenging due to the harsh environment, requiring specialized equipment, divers, or expensive remotely operated vehicles (ROVs).
One major hurdle is the low visibility caused by how light behaves in water. Water absorbs longer wavelengths quickly, leading to discolored footage. Natural light doesn’t penetrate deep, making artificial lighting necessary, which in turn can amplify backscattering and introduce issues like uneven illumination. The presence of suspended particles and sediment, known as turbidity, further intensifies light attenuation.
Because of these difficulties and high costs, real underwater data is scarce, often skewed towards clearer habitats, leaving murky environments largely unstudied. To address this data scarcity, synthetic data generation has emerged as a promising solution, especially for image enhancement and restoration where reference images are crucial. A common method involves using the underwater image formation model (IFM) to simulate various underwater effects on clear images. While computationally efficient, this method often produces unrealistic results, particularly in turbid conditions.
A recent research paper, “Sea-ing Through Scattered Rays: Revisiting the Image Formation Model for Realistic Underwater Image Generation,” explores these challenges and proposes significant improvements to synthetic underwater data generation. The authors, Vasiliki Ismiroglou, Malte Pedersen, Stefan H. Bengtson, Andreas Aakerberg, and Thomas B. Moeslund from Aalborg University, Denmark, highlight that many existing approaches focus on discoloration but overlook the model’s ability to capture complex, distance-dependent visibility loss in highly turbid settings. You can read the full paper here: Sea-ing Through Scattered Rays.
Rethinking the Underwater Image Formation Model
The core of this research lies in refining the underwater image formation model (IFM), also known as the Jaffe-McGlamery model. This model typically describes a degraded underwater image as a sum of three components: direct transmission (light directly from the scene), forward scattering (light scattered at small angles within the camera’s view, causing blur), and backscatter (light scattered towards the camera before reaching the scene, creating a hazy veil).
Historically, the forward scattering term has often been ignored due to perceived minimal effects and computational complexity. However, this paper argues that in turbid environments, blur from forward scattering is not negligible. The researchers propose an improved synthetic data generation pipeline that explicitly includes this commonly omitted forward scattering term. They also consider a non-uniform medium, which better reflects the chaotic fluid dynamics of real underwater environments.
Introducing the BUCKET Dataset
A significant challenge in synthetic image generation is the lack of ground truth data for evaluation. To tackle this, the researchers collected a new dataset called BUCKET (Baseline for Underwater Conditions in Known Environments and Turbidity). This dataset was gathered under controlled turbidity conditions in a cylindrical tank, using oat milk and clay to simulate varying levels of murkiness. Multiple GoPro cameras captured images of objects placed on a LEGO baseplate, ensuring consistent object positions between clear reference images and those collected as turbidity increased.
This controlled environment allows for accurate measurement of turbidity and provides corresponding reference images, making it invaluable for evaluating image restoration, object detection, and multi-view reconstruction tasks. The preliminary version of the BUCKET dataset, comprising 64 images captured under ambient light, will be made publicly available.
Qualitative Improvements and Survey Results
The proposed model demonstrates noticeable qualitative improvements over the standard IFM, especially under increasing turbidity. The inclusion of the forward scattering term helps capture the blur that becomes apparent even at short distances in murky conditions. Furthermore, the model introduces a stochastic approach to medium inhomogeneity using Gaussian Random Fields (GRFs). These GRFs help the simulated effects blend more naturally, creating smoother transitions and more realistic fading, while also mitigating artifacts from learning-based depth estimators.
To evaluate the realism of their synthetic images without direct references, the researchers conducted a survey with 20 computer vision experts. Participants were shown pairs of images—one generated by the reference model and one by the proposed model—and asked to choose the most realistic one. The results were compelling: the proposed pipeline was preferred in 73.9% of responses across all water types, and an impressive 82.5% for coastal water types, which are typically more turbid. This indicates a strong preference for the realism achieved by the new model, particularly in challenging high-turbidity scenarios.
Also Read:
- Evaluating YOLO Models for Underwater Object Detection: A Deep Dive into Performance and Efficiency
- Advanced AI Framework Boosts Ship Wake Detection in Radar Imagery
Looking Ahead
While the improvements are significant, the paper also highlights that accurate depth information remains a weak point for this approach to synthesizing data. The way depth scaling is altered can lead to large divergences in the resulting images, and the impact of inaccurate depth on downstream tasks needs further study.
In conclusion, this research offers a substantial step forward in generating realistic synthetic underwater data. By reintroducing the forward scattering term and accounting for medium inhomogeneity, the new pipeline produces visually superior results, especially in turbid conditions, which are crucial for advancing underwater computer vision applications.
