• 首页
  • 期刊简介
  • 编委会
  • 版权声明
  • 投稿指南
  • 期刊订阅
  • 相关下载
    雷达数据
    下载专区
  • 过刊浏览
  • 联系我们
引用本文:张丽丽, 王贤俊, 屈乐乐, 刘雨轩. SAR图像舰船检测的神经网络关联剪枝方法[J]. 雷达科学与技术, 2024, 22(3): 284-290.[点击复制]
ZHANG Lili, WANG Xianjun, QU Lele, LIU Yuxuan. Neural Network Associative Pruning for Ship Detection in SAR Images[J]. Radar Science and Technology, 2024, 22(3): 284-290.[点击复制]
【打印本页】   【下载PDF全文】   【查看/发表评论】  【下载PDF阅读器】  【关闭】
←前一篇|后一篇→ 过刊浏览    高级检索
本文已被:浏览 2344次   下载 333次 本文二维码信息
码上扫一扫!
分享到: 微信 更多
字体:加大+|默认|缩小-
SAR图像舰船检测的神经网络关联剪枝方法
张丽丽, 王贤俊, 屈乐乐, 刘雨轩
沈阳航空航天大学电子信息工程学院, 辽宁沈阳 110136
摘要:
合成孔径雷达(Synthetic Aperture Radar, SAR)由于其具备全天时、全天候的工作特点,使其在海洋环境监测、海洋资源调查和海洋防灾减灾等领域得到了广泛的应用。其中,基于SAR图像的舰船目标检测是SAR图像处理中的重要部分,其在军用和民用领域均具有重要的意义。本文针对基于深度学习实现的SAR图像目标检测算法参数计算量大、内存占用率高的问题提出了关联剪枝方法。该方法通过对网络进行改进,将相关联的卷积同时进行剪枝,并在训练结束后统一映射到低维度上以实现剪枝操作。通过在SSDD、SAR?Ship?Data?set和HRSID上进行实验,可以在保证平均精度(AP50)下降小于2%的前提下,针对FCOS网络实现70%以上的剪枝率,验证了所提方法的有效性。
关键词:  合成孔径雷达  模型剪枝  目标检测  深度学习
DOI:DOI:10.3969/j.issn.1672-2337.2024.03.006
分类号:TN959.72
基金项目:辽宁省教育厅资助项目(No.LJKZ0174,LJKMZ20220533)
Neural Network Associative Pruning for Ship Detection in SAR Images
ZHANG Lili, WANG Xianjun, QU Lele, LIU Yuxuan
College of Electronic Information Engineering, Shenyang Aerospace University, Shenyang 110136, China
Abstract:
The all?day and all?weather working characteristics of synthetic aperture radar (SAR) have made it widely used in the fields of marine environment monitoring, marine resources survey and marine disaster prevention and mitigation. Ship target detection based on SAR images is a critical element of SAR image processing, which is of great significance in both military and civilian fields. In this paper, a deep learning based associative pruning method is proposed to solve the problem of large parameter computation and high memory consumption of SAR image target detection algorithm. By improving the network, the method prunes the associated convolutions simultaneously, and maps them to the lower dimension after the training to realize the pruning operation. Through experiments on SSDD, SAR?Ship?Data?set and HRSID, it is possible to achieve a pruning rate to more than 70% for FCOS networks under the premise of ensuring that the average precision (AP50) decreases by less than 2%, which verifies the effectiveness of the proposed method.
Key words:  synthetic aperture radar(SAR)  model pruning  object detection  deep learning

版权所有:《雷达科学与技术》编辑部 备案:XXXXXXX
主办:中国电子科技集团公司第三十八研究所 地址:安徽省合肥市高新区香樟大道199号 邮政编码:230088
电话:0551-65391270 电子邮箱:radarst@163.com
技术支持:北京勤云科技发展有限公司