基于贝叶斯框架的全卷积网络和显式边缘网络修正的SAR图像分类

发布时间：2018-03-30 00:02

本文选题：合成孔径雷达　切入点：后置条件随机场　出处：《武汉大学》2017年硕士论文

【摘要】：合成孔径雷达(Synthetic Aperture Radar,SAR)信号相较于传统光学图像信号具有高解析度,多通道的特点,主要用于民用的如农作物监测,灾害监测,军用的如地图绘制,军事侦察等。一般将SAR信号转化为对应多通道图像,利用现有光学图像的方法来进行场景分类与标注,本文利用ESAR的Pauli分解伪彩图,首先用传统的方法来分类,然后使用神经网络和边缘约束的方法,相较于以前的方法提高了正确率,改善了边缘分布。在此之前经典的传统方法要进行超像素块的分割,块特征的提取,以及特征分类器这三个步骤,均值平移(meanshift)或分水岭方法常用于超像素块的提取,分类器多采用支持向量机(Support Vector Machine,SVM)或多层网络(Back Propagation Network,BP)等,之后也引入了马尔科夫场(Markov Random Filed,MRF)和条件随机场(Conditional Random Field,CRF)进一步考虑周围超像素块的分类类别从而进一步提高了正确率。首先,以前的方法的分类器多使用传统如SVM,BP,Logistic classifier等分类器,这些分类器并不适合图像信号,为提高图像初始的分类正确率,引入全卷积网络(Fully Convolutional Network,FCN)网络中效果最好的FCN8作为分类器,FCN8同时也负责卷积特征提取,同时FCN不需要用其他方法如meanshift来先将图像分块,再去针对区域处理,而是直接作用于单个像素点的RGB向量上。然后,为了考虑周围像素分类类别的影响,同时考虑自身像素的分类类别,引入Deeplab的全连接CRF综合考虑周围像素的类别先验信息,并用网格优化法得到合适的CRF相关参数。FCN网络用于分割时边缘部分的像素点分类不准确,这是反卷积和卷积的卷积核尺度决定的,在Deeplab内部利用区域变换(DomainTransform,DT)网络可以稍微修正边缘像素点的分类结果,但效果不显著,原DT算法存在很多问题,加以分析改进提出自己的改进DT扩散模型,在Deeplab外部利用FCN边缘网络提取边缘信息,分别与DT的内部边缘网络输出融合后输入DT中,然后与Deeplab的结果进行外部融合,融合方法包括我提出的改进DT的扩散法和补洞法。同时在内部和外部进行边缘融合最终可以将分类结果边缘改善以及提高一定的正确率。最后在ESAR数据上进行了多方案的对比实验,分别为基于potts先验的标准CRF或SVM分类模型并采取meanshift分割的方法,FCN8,Deeplab以及Deeplab-DT和最终的Deeplab-DT加FCN边缘网络融合的方法,实验表明每一种方案都比较以前的方案有一定提升,最终的边缘后置融合的方案效果最好,正确率最高同时边缘修正的效果很好。
[Abstract]:Compared with traditional optical image signals, synthetic Aperture radar synthetic Aperture radar (SAR) signals have the characteristics of high resolution and multi-channel. They are mainly used for civilian applications such as crop monitoring, disaster monitoring and military mapping. Military reconnaissance and so on. The SAR signal is generally converted into the corresponding multi-channel image, and the existing methods of optical image are used to classify and label the scene. In this paper, the pseudo-color image is decomposed by Pauli of ESAR. Firstly, the traditional method is used to classify the scene. Then, the neural network and edge constraint are used to improve the accuracy and edge distribution compared with the previous methods. Before this, the classical traditional methods are used to segment the super-pixel blocks and extract the features of the blocks. The three steps of feature classifier, mean shift mean shift (mean shift) or watershed method are often used to extract super-pixel blocks. Support vector machine support Vector machine (SVM) or multilayer network back Propagation network (BP) are often used in classifier. After that, Markov Random filter (MRFs) and conditional conditional Random (CRF) are introduced to further consider the classification classification of the surrounding super-pixel blocks, thus further improving the accuracy. Firstly, the traditional classifiers such as SVMBPPU Logistic classifier are often used in the former methods. These classifiers are not suitable for image signals. In order to improve the initial classification accuracy of images, FCN8, which is the most effective classifier in Fully Convolutional Network (FCN8) network, is also used to extract convolutional features. At the same time, FCN does not need to use other methods such as meanshift to divide the image into blocks first, then to deal with the region. Instead, it acts directly on the RGB vector of a single pixel, and then, in order to consider the effect of the classification of the surrounding pixels, At the same time, considering the classification categories of their own pixels, the fully connected CRF of Deeplab is introduced to synthetically consider the prior information of the surrounding pixels, and the appropriate CRF correlation parameter .FCN network is obtained by using the mesh optimization method to classify the pixels in the edge parts. This is determined by the convolution kernel scale of deconvolution and convolution. The classification results of edge pixels can be slightly modified by using domain transform DTN in Deeplab, but the effect is not significant. There are many problems in the original DT algorithm. The improved DT diffusion model is analyzed and improved. The edge information is extracted by using FCN edge network outside Deeplab, and then input into DT after output fusion with DT's inner edge network, and then external fusion is carried out with the result of Deeplab. The fusion methods include the improved DT diffusion method and the hole filling method proposed by me. At the same time, the edge fusion can improve the edge of classification results and improve the accuracy of the classification results. Finally, it is carried out on the ESAR data. A multi-program comparative experiment, It is a standard CRF or SVM classification model based on potts priori, and uses meanshift segmentation method to merge FCN8 and Deeplab, and Deeplab-DT and the final Deeplab-DT plus FCN edge network fusion method. The experiments show that each scheme has a certain improvement compared with the previous scheme. The final edge post fusion scheme has the best effect and the highest correct rate and good edge correction effect.
【学位授予单位】：武汉大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN957.52

【相似文献】