Physics-Informed Synthetic Underwater Benchmark Dataset for Image Enhancement
A new framework called π-SUB has been developed by researchers to create synthetic underwater benchmark datasets, targeting the reduction of the synthetic-to-real gap in Underwater Image Enhancement (UIE). This framework enhances the traditional underwater image formation model by integrating depth-dependent downwelling irradiance, biologically resolved absorption, and environmental scattering for all ten Jerlov water types, while allowing for independent control of residual effects. The dataset generated includes paired synthetic underwater-reference images that cover a range from shallow to deep and coastal to oceanic settings. Comprehensive simulation studies assessed π-SUB's hyper-realism, achieving a global Fréchet Inception Distance (FID) that is 46% lower than Syrea, and its generalizability was tested with four leading UIE architectures. The research can be found on arXiv under ID 2608.10589.
Key facts
- π-SUB is a physics-informed framework for generating synthetic underwater benchmark datasets.
- It extends the classical underwater image formation model with depth-dependent downwelling irradiance, biologically resolved absorption, and environmental scattering.
- The dataset covers all ten Jerlov water types and includes shallow-to-deep and coastal-to-oceanic environments.
- π-SUB achieves a global FID 46% lower than Syrea, indicating improved hyper-realism.
- Four state-of-the-art UIE architectures were used to evaluate generalizability.
- The paper is published on arXiv with ID 2608.10589.
- The framework aims to bridge the synthetic-to-real gap in Underwater Image Enhancement.
Entities
Institutions
- arXiv