基于深度学习的SAR目标识别方法研究
发布时间：2019-04-19 22:01
【摘要】：合成孔径雷达(Synthetic Aperture Radar,SAR)由于其独特的优势已经成为当今社会的一种重要的信息获取手段,无论在军用领域还是民用领域都发挥着至关重要的作用。作为获取SAR信息的方式,SAR图像的识别一直是研究热点之一。近年来深度学习的提出引起了又一股人工智能的研究热,深度学习由于将非监督学习和监督学习结合,使得大量的无标签的数据都有了学习的价值,因而在目标识别方面取得了前所未有的成功,但仍面临着许多问题。本文首先总结了基于机器学习的SAR图像目标识别的技术,给出了监督学习中的神经网络和非监督学习中的主成分分析两种方法在MSTAR数据集上的识别效果;其次本文指出了机器学习方法在目标识别方面的局限,即:在应用到SAR图像目标识别时需要大量的专业知识,不能自动的提取能够表征SAR目标的特征,基于此,本文提出了深度学习的模型可以解决该问题,分别将深度置信网络和卷积神经网络两种深度学习模型用于SAR图像目标识别,并分析了两种模型的各个参数对模型的性能的影响,给出了在用于识别SAR目标时的这些参数的典型值;最后,由于SAR图像中包含大量的相干斑噪声,这是影响模型识别性能的关键因素之一,本文在对比了Lee滤波和小波变换两种相干斑噪声抑制方法的识别效果的基础上,得出结论:Lee滤波和二层小波变换两种方法的结合可以获得在识别性能方面的提升。
[Abstract]:Because of its unique advantages, synthetic aperture radar (Synthetic Aperture Radar,SAR) has become an important means of information acquisition in today's society. It plays an important role in both military and civilian fields. As a way to obtain SAR information, SAR image recognition has always been one of the research hotspots. In recent years, the proposal of in-depth learning has caused another hot research in artificial intelligence. Because of the combination of unsupervised learning and supervised learning, a large number of unlabeled data have the value of learning. As a result, it has achieved unprecedented success in target recognition, but it still faces many problems. In this paper, the technology of SAR image target recognition based on machine learning is summarized, and the recognition effects of neural network in supervised learning and principal component analysis in unsupervised learning on MSTAR data sets are given. Secondly, this paper points out the limitation of machine learning method in target recognition, that is, when it is applied to SAR image target recognition, it needs a lot of professional knowledge, and can not automatically extract the features of SAR target, which is based on this. In this paper, a depth learning model is proposed to solve this problem. Two depth learning models, depth confidence network and convolution neural network, are applied to target recognition of SAR images, and the influence of each parameter of the two models on the performance of the model is analyzed. The typical values of these parameters used to identify SAR targets are given. Finally, because there is a lot of speckle noise in the SAR image, which is one of the key factors affecting the performance of model recognition, this paper compares the recognition effects of two coherent speckle suppression methods, Lee filter and wavelet transform, based on the comparison of the recognition results of the two coherent speckle suppression methods, Lee filtering and wavelet transform. It is concluded that the combination of Lee filtering and bilevel wavelet transform can improve the recognition performance.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN957.52
本文编号：2461327
[Abstract]:Because of its unique advantages, synthetic aperture radar (Synthetic Aperture Radar,SAR) has become an important means of information acquisition in today's society. It plays an important role in both military and civilian fields. As a way to obtain SAR information, SAR image recognition has always been one of the research hotspots. In recent years, the proposal of in-depth learning has caused another hot research in artificial intelligence. Because of the combination of unsupervised learning and supervised learning, a large number of unlabeled data have the value of learning. As a result, it has achieved unprecedented success in target recognition, but it still faces many problems. In this paper, the technology of SAR image target recognition based on machine learning is summarized, and the recognition effects of neural network in supervised learning and principal component analysis in unsupervised learning on MSTAR data sets are given. Secondly, this paper points out the limitation of machine learning method in target recognition, that is, when it is applied to SAR image target recognition, it needs a lot of professional knowledge, and can not automatically extract the features of SAR target, which is based on this. In this paper, a depth learning model is proposed to solve this problem. Two depth learning models, depth confidence network and convolution neural network, are applied to target recognition of SAR images, and the influence of each parameter of the two models on the performance of the model is analyzed. The typical values of these parameters used to identify SAR targets are given. Finally, because there is a lot of speckle noise in the SAR image, which is one of the key factors affecting the performance of model recognition, this paper compares the recognition effects of two coherent speckle suppression methods, Lee filter and wavelet transform, based on the comparison of the recognition results of the two coherent speckle suppression methods, Lee filtering and wavelet transform. It is concluded that the combination of Lee filtering and bilevel wavelet transform can improve the recognition performance.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN957.52
【参考文献】
相关期刊论文 前4条
1 王雅思;姚鸿勋;孙晓帅;许鹏飞;赵思成;;深度学习中的自编码器的表达能力研究[J];计算机科学;2015年09期
2 刘建伟;刘媛;罗雄麟;;深度学习研究进展[J];计算机应用研究;2014年07期
3 贾承丽;赵凌君;吴其昌;匡纲要;;基于遗传算法的SAR图像道路网检测方法[J];计算机学报;2007年07期
4 程辉;于秋则;田金文;柳健;;基于小波支持向量机分割的SAR图像桥梁目标检测[J];华中科技大学学报(自然科学版);2006年04期
，本文编号：2461327
本文链接：https://www.wllwen.com/kejilunwen/xinxigongchenglunwen/2461327.html
