摘要
深度学习为遥感领域诸多应用提供了重要的技术支撑,光学遥感图像的舰船目标检测对国防侦察和预警具有重要意义。真实场景中的舰船往往呈不同方向任意排列,且小目标的占比大,经典的深度学习目标检测算法在这种复杂条件下精度低、易漏检。为此,本文设计了基于注意力机制特征重建网络的舰船目标检测算法。首先,通过引入注意力机制对多尺度特征融合网络模型进行训练,以高召回率产生水平锚框;然后,旋转锚框以缓解密集排列目标引起的噪声问题,并利用特征重建模块来缓解特征不对齐的问题,实现模型精炼。在HRSC2016和DOTA数据集上的测试结果表明:舰船目标检测平均精度分别达到90.20和87.52,相比经典的深度学习目标检测算法得到了有效提升,并在模拟星载嵌入式智能图像处理平台上验证了算法在轨应用的可行性。
Deep learning provides essential supports to various applications in the remote sensing field. Ship detection from optical remote sensing images has great significance for national defense reconnaissance and early warning. Ships in real scenes usually orient arbitrarily,and most of them are of tiny size. Classical detection algorithms based on deep learning have low accuracy and high miss rate under such complex conditions. Therefore,a ship detection algorithm based on the reconstruction network with attention mechanism features is proposed in this paper.The horizontal anchor boxes with high recall rate are generated by introducing the attention mechanism to train the multi-scale feature fusion network. Subsequently,the noise and feature misalignment are alleviated to refine the network model by rotating the anchor box and employing the feature reconstruction module successively. The ship detection results testing on HRSC2016 and DOTA data sets show that the average accuracy achieved with the proposed algorithm is 90.20 and 87.52,respectively,which are improved effectively compared with those obtained with the classical deep learning based algorithms. Besides,the feasibility of in-orbit application is verified on a simulated satellite-borne embedded intelligent image processing platform.
作者
牛戈
陈小前
季明江
郭鹏宇
刘勇
冉德超
NIU Ge;CHEN Xiaoqian;JI Mingjiang;GUO Pengyu;LIU Yong;RAN Dechao(National Innovation Institute of Defense Technology,Academy of Military Sciences,Beijing 100071,China)
出处
《上海航天（中英文）》
CSCD
2021年第4期128-136,共9页
Aerospace Shanghai（Chinese&English）
基金
国家自然科学基金(61901504)。
关键词
遥感图像
舰船检测
小目标
深度学习
注意力机制
remote sensing image
ship detection
small object
deep learning
attention mechanism
