Title: Lightweight SAR Ship Detection via Contrastive Distillation

Abstract:

While deep convolutional and transformer-based detectors deliver robust results for Synthetic Aperture Radar (SAR) ship detection, their high computational demands often hinder real-time or onboard implementation. Although lightweight models enhance efficiency, they frequently fail to capture the intricate structural relationships present in SAR backscatter data. Current knowledge-distillation techniques for SAR primarily depend on feature or logit matching, a method that enforces similarity in localized activations but overlooks the geometric dependencies among object representations.

To address these limitations, we introduce SURGE (Structured Unified Relational knowledGE distillation framework for SAR Ship detection). This approach transfers relational geometry from a high-performance teacher detector to a streamlined student detector by employing a contrastive InfoNCE objective within a shared projection embedding space. To the best of our knowledge, this represents the first knowledge-distillation framework for transformer-based SAR ship detectors. The proposed framework is architecture-agnostic, offering a universal region-level distillation interface compatible with two-stage, one-stage, and transformer-based detectors, thereby requiring no modifications to their existing deployed structures. Evaluations on the SSDD and HRSID benchmarks reveal that the method significantly boosts performance for two-stage detectors, delivering improvements of up to 6.2 mAP and 8.0 AP75 over baseline student models, and even outperforming the teacher model itself.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC