超轻量化SAR影像小目标检测网络

打开文本图片集

关键词：SAR影像;小目标检测；超轻量化；多分支高效层聚合;增强共享检测头；剪枝与知识蒸馏中图分类号：TP751.1文献标识码：Adoi：10.37188/OPE.20253310.1672 CSTR：32169.14.OPE.20253310.1672

Ultra lightweight SAR image small object detection network

YANG Xiaomin 1，3，4 ， YANG Jun1，2，3，4*

（1. Faculty of Geomatics， Lanzhou Jiaotong University， Lanzhou 73oO7O， China;

2. School of Electronic and Information Engineering， Lanzhou Jiaotong University， Lanzhou 730070，China;

3. National and Local Joint Engineering Research Center of Geographical Monitoring Technology Application，Lanzhou 73007O， China;

4. Gansu Prouincial Engineering Laboratory of Geographical Monitoring，Lanzhou 730o7O， China） * Corresponding author， E-mail： yangj@mail. lzjtu. cn

Abstract： Although the synthetic aperture radar （SAR） image target detection method utilizing convolu tional neural network technology can achieve good detection accuracy，its high model complexity limits its practical application and deployment in military rapid decision-making，maritime emergency rescue，and other fields. Therefore，this paper proposes an ultra lightweight smalltarget detection model for radar images. Firstly，a multi branch eficient layer aggregation module is designed to enhance multi-scale perception and adapt to various resources and computing capabilities of actual devices. Secondly，detail enhancementand shared detection heads are utilized to focus on smalltarget information，thereby reducing false detections caused by sea and land clutter interference.Finally，feature richnes-guided pruning and knowledge distilation guided representation learning are employed to further compress the model and enhance performance. The experimental results demonstrate that the network model achieves detection accuracies of 89.0% ， 98.1% ， 82.5% ， 98.6% ，and 91.5% on MSAR，SAR-Ship，AIR-SARShip-2.0，SSDD， and HRSID datasets，respectively，with a computational complexity of 4.186 G and a parameter complexity of 0.888M . The algorithm presented in this paper exhibits good robustness，and the network model can achieve optimal detection speed and accuracy at the minimum volume.

Key words： SAR image； small object detection； ultra lightweight;multi branch eficient layer aggregation；enhance shared detection head；pruning and knowledge distillation

1引言

合成孔径雷达（SAR）是一种通过主动发射和接收电磁波实现微波成像的传感器，具备穿透云层、雨雪等天气条件的能力，能够实现全天候、全天时的对地观测，因此在灾害应急、空间侦察和环境监测等领域具有重要应用。（剩余24524字）