SURGE framework boosts lightweight SAR ship detection by 8 AP75 points
New contrastive distillation method lets tiny models match giant teachers in radar imagery.
SAR ship detection is critical for maritime surveillance but deep models are too heavy for onboard deployment. Existing lightweight models lose structural relationships in radar backscatter, and prior knowledge distillation methods only match features or logits. To bridge this, researchers introduce SURGE—a Structured Unified Relational knowledGE distillation framework that transfers geometric relationships between object representations from a powerful teacher to a compact student. Using a contrastive InfoNCE objective in a shared embedding space, SURGE captures the complex spatial structure of ship targets missed by simpler distillation. It's architecture-agnostic, providing a common region-level distillation interface for one-stage, two-stage, and transformer-based detectors without requiring any architectural changes.
On the SSDD and HRSID benchmarks, SURGE delivers significant gains: up to 6.2 mAP and 8.0 AP75 over baseline students, and in some cases the student even exceeds the teacher's performance. This is the first transformer-based knowledge distillation framework for SAR ship detection. The work was accepted at GLSVLSI'26 and could enable real-time, high-accuracy ship detection on edge devices like drones and satellites, where compute is limited but detection precision is vital.
- SURGE uses contrastive InfoNCE loss to transfer relational geometry from teacher to student detectors.
- Achieves up to 6.2 mAP and 8.0 AP75 improvement over lightweight baselines on SSDD and HRSID.
- Architecture-agnostic: works with two-stage, one-stage, and transformer-based detectors without modification.
Why It Matters
Enables accurate ship detection on edge devices like satellites and drones, critical for real-time maritime surveillance.