基于更新学习机制的SAR图像目标识别方法研究

发布时间：2019-04-30 13:36

【摘要】：合成孔径雷达(Synthetic Aperture Radar,SAR)具有广阔的研究和应用前景,其被广泛地应用于包含地质勘测、地形测绘和制图、海洋应用在内的民用领域。而在军事领域中,SAR可用作全天时全天候全球战略侦察与重点战区军事监控。SAR目标识别的研究主要依赖于大规模的带标签数据样本,然而,要获得大量的样本数据及其正确的分类标签,需要耗费非常大的人力物力资源。不仅如此,基于庞大数据训练分类器同样耗费大量资源并且难度较高。因此,如何在保证识别性能的前提下降低资源开销成为了一个新的热点研究方向。通过较少训练样本得到的分类器性能不够成熟,因此SAR图像数量的稀少,是制约SAR图像目标识别发展的主要原因之一。然而,SAR图像样本数量将随着时间的推移不断增长,同时,又希望能够直接利用新增的SAR图像提升原有分类器性能,而不需要与原有的图形样本混合重新训练新分类器,以此达到减小训练开支的目的.将利用不断新增的样本在已有的分类器基础上,迭代训练以提升分类器性能的方法定义为更新学习机制。本文研究主要针对于SAR图像目标识别问题,研究基于卷积神经网络的SAR图像特征提取方法,并建立一种基于更新学习机制的SAR图像目标识别方法。本文的主要内容分为以下部分:1、介绍SAR图像目标识别的背景,分析该课题的研究意义,以及国内外学者在SAR图像目标识别领域所做的主要研究;2、分析SAR图像特性及其特征提取方法,引入机器学习理论于MSTAR数据库;3、实现基于支持向量机的SAR图像目标识别方法,并重点介绍深度学习在目标识别领域的应用,利用深度学习中的卷积神经网络理论实现SAR图像目标识别,并与其他方法进行对比;4、以卷积神经网络为主,支持向量机为辅助分类器实现基于卷积神经网络与SVM的更新学习机制。
[Abstract]:Synthetic Aperture Radar (Synthetic Aperture Radar,SAR) has a wide range of research and application prospects. It has been widely used in civil fields, including geological survey, topographic mapping and mapping, and marine applications. In the military field, SAR can be used as global strategic reconnaissance and military surveillance in all-weather and all-weather. The research of target recognition mainly depends on large-scale labeled data samples, however, the research of target recognition is based on large-scale labeled data samples. In order to obtain a large amount of sample data and correct classification label, it takes a lot of human and material resources. Moreover, training classifiers based on huge data also costs a lot of resources and is difficult. Therefore, how to reduce resource overhead under the premise of ensuring recognition performance has become a new hot research direction. The performance of classifier obtained by fewer training samples is not mature enough, so the scarcity of SAR images is one of the main reasons restricting the development of SAR image target recognition. However, the number of SAR image samples will continue to increase with the passage of time. At the same time, it is hoped that the performance of the existing classifier can be improved directly by using the new SAR image, without the need to re-train the new classifier with the existing graphics samples. In order to achieve the purpose of reducing the cost of training. The updated learning mechanism is defined as the method of iterative training to improve the performance of classifiers based on the existing classifiers. This paper focuses on the problem of target recognition for SAR images, studies the feature extraction method of SAR images based on convolution neural network, and establishes a SAR image target recognition method based on update learning mechanism. The main contents of this paper are as follows: 1. The background of SAR image target recognition is introduced, and the research significance of this subject is analyzed, as well as the main research done by domestic and foreign scholars in the field of SAR image target recognition; 2. The characteristics of SAR image and its feature extraction methods are analyzed, and the machine learning theory is introduced into the MSTAR database. 3, the method of SAR image target recognition based on support vector machine is realized, and the application of depth learning in target recognition field is introduced emphatically. The convolutional neural network theory of depth learning is used to realize SAR image target recognition. And compared with other methods; 4, based on convolution neural network and support vector machine as assistant classifier, the updating learning mechanism based on convolution neural network and SVM is realized.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN957.52

【参考文献】