A Lightweight Framework With Pruning and Knowledge Distillation for Ship Detection in SAR Images

Lightweight synthetic aperture radar (SAR) ship detection is critical for real-time maritime monitoring on resource-constrained platforms. However, speckle noise and sea clutter often obscure small and low-contrast vessels, making accurate detection difficult under strict computational budgets. To address this, we propose a compact detection framework that integrates layerwise amplitude-based pruning with dual-stream knowledge distillation, enabling substantial model compression while maintaining detection performance. In addition, we adopt efficiency-oriented building blocks to improve feature utilization and introduce a localization-aware regression objective to strengthen bounding box fitting on low intersection over union (IoU) samples. Experiments on three public SAR ship detection benchmarks demonstrate that the proposed pruned and distilled model achieves higher precision and mean average precision (mAP) than recent peer methods, while running above 200 frames per second (FPS) with only 7.4 M parameters and 41.9 G floating point operations (FLOPs). These results indicate that pruning-guided distillation provides an effective and practical path toward deployable SAR ship detection.