面向SAR图像目标识别和地物分类的深度学习研究

发布时间：2018-05-06 06:20

本文选题：DBN + SAR　；参考：《西安电子科技大学》2015年硕士论文

【摘要】：深度学习起源于人工神经网络,模仿人脑计算模式,可以自动地分层学习出抽象特征,在图像领域应用广泛,尤其是在目标识别和图像分类方面。随着遥感技术的发展,合成孔径雷达(Synthetic aperture radar,SAR)图像以其大信息量,全天候全天时的特点在军事和民用领域占据了重要地位。对于SAR图像的识别和分类任务而言,选取合适的特征非常重要,所用特征决定了算法性能的上限,而深度学习模型可以自动从原始数据中学出更抽象的特征。在深度学习模型中,深度置信网络(Deep Belief Network,DBN)使用的栈式限制玻尔兹曼机(Restricted Boltzmann Machine,RBM)的非监督学习和反向传播的有监督微调过程,可以自动学习到更适合分类的特征。本文用深度置信网络来提取高层抽象特征用于SAR图像。具体工作如下:一,用于单极化SAR目标识别的深度学习研究。由于深度学习更适用于大数据,而本文所用的运动与静止目标的获取与识别数据(MSTAR)数量有限,使得深度学习模型不容易收敛,所以我们提出数据融合与深度学习相结合的策略。分别提取MSTAR的轮廓波特征和曲线波特征与原数据相结合作为深度置信网络的输入,同时加入更能模拟数据的高斯限制玻尔兹曼机(gaussianRBM),进行SAR图像的目标识别,识别精度较原始RBM和单一数据有所提高。由于传统DBN没有考虑到SAR图像的2-D结构和空间信息,导致学习到的权值与像素所处位置无关,而卷积网络的权值共享使得每一种权值对应一种特征算子,更利于提取不同性质的特征,所以,本文使用基于卷积RBM的深度置信网络,使得识别精度进一步提高。二,用于全极化SAR图像地物分类的深度学习研究。由于传统的RBM更适合模拟二值数据,对于符合其他指数家族的分布,RBM可以加入不同的统计特性进行扩展。所以,对极化SAR实数数据,我们使用加入高斯分布的gaussianRBM构成DBN,用于极化SAR图像的地物分类;对极化SAR复数数据,我们基于极化SAR数据复wishart分布特性提出了的wishartRBM,并由此构成DBN,用于极化SAR图像的地物分类。具体步骤为:将极化SAR数据的协方差矩阵元素作为输入,先用多层的wishartRBM预训练网络,再加上反向传播进行微调,最后使用softmax分类进行地物分类,分类精度与其他方法相比得到了提高。
[Abstract]:Depth learning originates from artificial neural network and imitates the human brain computing model. It can automatically learn abstract features in layers and is widely used in image field, especially in target recognition and image classification. With the development of remote sensing technology, synthetic aperture radar synthetic aperture radar (SAR) images play an important role in military and civilian fields with its large amount of information and all-weather and all-day characteristics. For the task of SAR image recognition and classification, it is very important to select suitable features, which determine the upper bound of the algorithm performance, while the depth learning model can automatically learn more abstract features from the original data. In the deep learning model, the unsupervised learning and backpropagation process of unsupervised learning and backpropagation of the deep confidence network Deep Belief Network (DBN) can automatically learn more suitable features for classification. In this paper, a depth confidence network is used to extract high-level abstract features for SAR images. The main work is as follows: 1. The research of deep learning for single polarization SAR target recognition. Because depth learning is more suitable for big data, and the number of moving and static target acquisition and recognition data is limited in this paper, it is difficult to converge in depth learning model, so we propose a combination of data fusion and depth learning strategy. The contour wave feature, curve wave feature and original data of MSTAR are extracted respectively as input of depth confidence network, and Gao Si restricted Boltzmann machine, which can simulate data, is added to realize target recognition of SAR image. The recognition accuracy is improved compared with the original RBM and single data. Because the traditional DBN does not consider the 2-D structure and spatial information of the SAR image, the weights learned are independent of the location of the pixels, and the weights of the convolutional network share the weights so that each weight corresponds to a feature operator. Therefore, the depth confidence network based on convolution RBM is used to improve the recognition accuracy. Secondly, it is used to study the depth learning of ground object classification in fully polarized SAR images. Because the traditional RBM is more suitable for simulating binary data, different statistical properties can be added to the distributed RBM which accords with other exponential families. So, for the real data of polarimetric SAR, we use the gaussianRBM with Gao Si distribution to form DBNs, which is used to classify the ground objects in polarized SAR images, and for the complex data of polarimetric SAR, Based on the complex wishart distribution of polarized SAR data, we propose wishart RBM, which is used to classify ground objects in polarimetric SAR images. The concrete steps are as follows: the element of covariance matrix of polarized SAR data is taken as input, the multi-layer wishartRBM pre-training network is used first, then the backpropagation is used to fine tune, and finally, softmax classification is used to classify ground objects. The classification accuracy is improved compared with other methods.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN957.52

【参考文献】