OSCAR
KNU PRMI Lab
Overview
OSCAR (Optical-aware Semantic Control for Aleatoric Refinement) is a novel framework for converting Synthetic Aperture Radar (SAR) imagery into photorealistic optical images. This work addresses the fundamentally ill-posed problem of SAR-to-optical translation.
Key Challenge
Translating SAR observations into photo-realistic optical images is fundamentally ill-posed due to speckle noise, geometric distortions, semantic misinterpretation, ambiguous texture synthesis, and structural hallucinations.
Translating SAR observations into photo-realistic optical images is fundamentally ill-posed due to speckle noise, geometric distortions, semantic misinterpretation, ambiguous texture synthesis, and structural hallucinations.
Approach
OSCAR integrates three core innovations:
| Component | Description |
|---|---|
| Cross-Modal Semantic Alignment | Establishes an Optical-Aware SAR Encoder by transferring robust semantic knowledge from an optical reference model to the SAR processor |
| Semantically-Grounded Generative Guidance | Implements a specialized ControlNet combining class-aware text prompts for broader context with hierarchical visual prompts for local spatial direction |
| Uncertainty-Aware Objective | Explicitly models aleatoric uncertainty to dynamically adjust reconstruction focus, addressing speckle-induced ambiguity challenges |
Key Results
OSCAR achieves superior perceptual quality and semantic consistency compared to state-of-the-art approaches through experimental validation.
Project Site
Resources
