基于层自适应稀疏剪枝策略的高效SAR舰船目标检测方法研究

Research on an Efficient SAR Ship Target Detection Method Based on Layer-Adaptive Sparse Pruning Strategy

摘要: 合成孔径雷达(SAR)舰船检测在海洋监视中至关重要，但现有的高精度模型(如YOLO11)参数量大、计算复杂，难以部署在星载或边缘计算等资源受限平台。针对上述问题，文章提出了一种面向SAR任务的YOLO11结构化剪枝框架SASP (SAR-Aware Structured Pruning)。该框架对原始模型进行了深入的架构感知冗余性分析，量化了C3k2模块内部及跨层连接中的特征相关性，精准定位了冗余结构。通过对基准模型进行多梯度的通道剪枝与充分微调，实验结果表明适度的剪枝操作能有效去除网络冗余特征，产生显著的正则化效应。该方法不仅大幅降低了计算成本，更实现了检测精度的逆势增长：在2.5加速比下，达到了89.9% mAP50高精度的同时实现了近1000 FPS的实时推理速度。研究证明了结构化剪枝在特定策略下可作为一种高效的正则化手段，为资源受限场景下的实时高精度目标检测提供了新的解决方案。

Abstract: Synthetic Aperture Radar (SAR) ship detection is crucial in maritime surveillance, but existing high-precision models (such as YOLO11) have large parameter counts and computational complexity, making them difficult to deploy on resource-constrained platforms such as spaceborne or edge computing. To address these issues, this paper proposes a YOLO11 structured pruning framework, SASP (SAR-Aware Structured Pruning), for SAR tasks. This framework performs in-depth architecture-aware redundancy analysis on the original model, quantifies the feature correlations within the C3k2 module and across layer connections, and accurately locates redundant structures. Through multi-gradient channel pruning and sufficient fine-tuning of the benchmark model, experimental results show that appropriate pruning operations can effectively remove network redundant features and produce a significant regularization effect. This method not only significantly reduces computational costs but also achieves a counter-trend increase in detection accuracy: at a speedup ratio of 2.5, it achieves a high accuracy of 89.9% mAP50 while realizing a real-time inference speed of nearly 1000 FPS. This study demonstrates that structured pruning, under specific strategies, can serve as an efficient regularization method, providing a new solution for real-time, high-precision target detection in resource-constrained scenarios.