李健伟^1,,,
曲长文¹,
彭书娟¹,
江源²
1.海军航空大学 ??烟台 ??264001
2.92228部队 ??北京 ??100044
基金项目:国家自然科学基金(61571454)

详细信息

作者简介:李健伟：男，1989年生，博士生，研究方向为SAR图像处理、机器学习及深度学习
曲长文：男，1964年生，教授，研究方向为雷达信号处理，信息对抗，信号与信息处理等
彭书娟：女，1980年生，博士生，研究方向为SAR图像处理

通讯作者:李健伟　lgm_jw@163.com

中图分类号:TN957.51

计量

文章访问数:1034
HTML全文浏览量:454
PDF下载量:98
被引次数:0

出版历程

收稿日期:2018-01-15
修回日期:2018-09-26
网络出版日期:2018-10-22
刊出日期:2019-01-01

Ship Detection in SAR images Based on Generative Adversarial Network and Online Hard Examples Mining

Jianwei LI^1,,,
Changwen QU¹,
Shujuan PENG¹,
Yuan JIANG²
1. Naval Aeronautical University, Yantai 264001, China
2. The Unit 92228 of PLA, Beijing 100044, China
Funds:The National Natural Science Foundation of China (61571454)


摘要
摘要:基于深度学习的SAR图像舰船目标检测算法对图像的数量和质量有很高的要求，而收集大体量的舰船SAR图像并制作相应的标签需要消耗大量的人力物力和财力。该文在现有SAR图像舰船目标检测数据集(SSDD)的基础上，针对目前检测算法对数据集利用不充分的问题，提出基于生成对抗网络(GAN)和线上难例挖掘(OHEM)的SAR图像舰船目标检测方法。利用空间变换网络在特征图上进行变换，生成不同尺寸和旋转角度的舰船样本的特征图，从而提高检测器对不同尺寸、旋转角度的舰船目标的适应性。利用OHEM在后向传播过程中发掘并充分利用难例样本，去掉检测算法中对样本正负比例的限制，提高对样本的利用率。通过在SSDD数据集上的实验证明以上两点改进对检测算法性能分别提升了1.3%和1.0%，二者结合提高了2.1%。以上两种方法不依赖于具体的检测算法，且只在训练时增加步骤，在测试时候不增加计算量，具有很强的通用性和实用性。
关键词:SAR图像/
舰船检测/
生成对抗网络/
线上难例挖掘
Abstract:Deep learning based ship detection method has a strict demand for the quantity and quality of the SAR image. It takes a lot of manpower and financial resources to collect the large volume of the image and make the corresponding label. In this paper, based on the existing SAR Ship Detection Dataset (SSDD), the problem of insufficient utilization of the dataset is solved. The algorithm is based on Generative Adversarial Network (GAN) and Online Hard Examples Mining (OHEM). The spatial transformation network is used to transform the feature map to generate the feature map of the ship samples with different sizes and rotation angles. This can improve the adaptability of the detector. OHEM is used to discover and make full use of the difficult sample in the process of backward propagation. The limit of positive and negative proportion of sample in the detection algorithm is removed, and the utilization ratio of the sample is improved. Experiments on the SSDD dataset prove that the above two improvements improve the performance of the detection algorithm by 1.3% and 1.0% respectively, and the combination of the two increases by 2.1%. The above two methods do not rely on the specific detection algorithm, only increase the time in training, and do not increase the amount of calculation in the test. It has very strong generality and practicability.
Key words:SAR images/
Ship detection/
Generative Adversarial Network (GAN)/
Online Hard Examples Mining (OHEM)



PDF全文下载地址:

https://jeit.ac.cn/article/exportPdf?id=aa0bd85f-84e3-4e32-a55b-39b3ba6a690e