基于深度卷积神经网络和无监督K均值特征的SAR图像目标识别方法研究

发布时间：2018-04-15 22:17

本文选题：合成孔径雷达 + 自动目标识别　；参考：《五邑大学》2017年硕士论文

【摘要】：图像识别是模式识别领域的典型应用,它通过计算机算法对图像进行处理、分析和理解,以识别图像中各种不同模式的对象。合成孔径雷达(Synthetic Aperture Radar,SAR)图像在当今科技的快速发展且日新月异的情况下,在各个领域尤其是军事领域发挥越来越多的作用,越来越多的学者将研究重心集中在SAR图像目标识别。而其中的自动目标识别(Automatic Target Recognition,ATR)技术是SAR图像识别系统的核心技术,由于ATR在对图像特征的学习以及提取过程所发挥的作用,使其成为国内外学者的研究热点。ATR主要借助于SAR图像数据,依靠计算机算法对图像目标区域进行定位,并从中提取出包含大量目标信息的图像特征,并对特征信息进行识别。近年来,很多SAR图像目标识别算法被提出,类似于其他图像识别算法的三个过程:图像预处理、图像目标特征提取和图像目标识别,本文同样从该三个方面出发,主要做了以下工作:(1)对于图像预处理,SAR图像中有大量的背景噪声以及图像目标占SAR图像比例较小,本文首先采用质心定位法寻找SAR图像感兴趣区域(Region of Interest,ROI),通过对原始图像感兴趣区域的提取以达到去除SAR图像背景噪声的效果。在本文的实验中,SAR原始图像中提取的ROI图像大小分别为49×49和64×64,实验结果表明ROI图像可以很好地去除SAR图像背景噪声;其次,在机器学习尤其是无监督学习和深度学习中,算法对训练数据的数据量有较高要求,同时SAR图像往往局限于图像数据量,因此,本文结合SAR图像成像时对目标方位角敏感的特点,提出了两种基于SAR图像的数据增强的方法,通过旋转图像中目标物体的方位角,以及在原始图像上增加随机整数值来得到更多图像数据,实验结果表明所提方法具有有效性。(2)深度学习(Deep Learning,DL)是机器学习中一大热门,在许多领域如图像识别、语音识别、自然语言处理等领域取得了突破。卷积神经网络(convolutional neural networks,CNN)是深度学习算法的一种,本文结合CNN原理,训练了一个深度CNN网络模型SARnet,SARnet包含两个卷积层,卷积核大小为7×7、两个池化层、两个全连接层;结合数据增强方法对训练集扩充为原始数据库的22倍之后,对网络进行训练,并用训练得到的网络模型对MSTAR数据库测试集进行特征提取,用SVM对提取出的特征进行分类,最终识别率达到了95.68%,该识别率高于其他CNN模型。(3)近来在机器学习中,研究学者已经将目光集中在从一些没有标签的数据中学习特征,本文针对SAR图像特征提取,先采用感兴趣区域提取使输入图像大小为64×64;然后采用无监督K均值特征学习算法,结合数据增强后的训练数据库,学习到一些有用的表示。通过分块自编码和优化接受域参数进行SAR图像特征学习可以使模型学习到多样性的特征,并通过实验证明,结合K-means的无监督特征学习得到的特征可以使SAR图像识别率达到96.67%的主流识别率。
[Abstract]:Image recognition is a typical application in the field of pattern recognition, which based on computer image processing, analysis and understanding, to identify images in different modes. The synthetic aperture radar (Synthetic Aperture, Radar, SAR) image in the rapid development of science and technology change rapidly and the situation in various fields, especially play a more and more the role of the military field, more and more scholars focused on SAR image target recognition. The automatic target recognition (Automatic Target Recognition ATR) technology is the core technology of SAR image recognition system don't, because ATR plays in the image feature extraction process and the learning effect, make it become a hot research topic.ATR at home and abroad with the help of SAR image data, relying on the target region of image positioning algorithm, and extract contains a large number of orders Image character information, and identification of feature information. In recent years, a lot of SAR image target recognition algorithm is proposed. The three process is similar to the other image recognition algorithm: image preprocessing, image feature extraction and image recognition, this paper also from the three aspects, mainly do the following work: (1) for image preprocessing, SAR images have a lot of background noise and image target accounted for a smaller proportion of SAR image, this paper uses centroid positioning method for SAR image region of interest (Region of, Interest, ROI), in order to remove the SAR image background noise effect on the original image by extracting the region of interest. In this experiment, the ROI image size extraction were 49 * 49 and 64 * 64 SAR in the original image. Experimental results show that the ROI image can remove the background noise of SAR image; secondly, especially in machine learning It is an unsupervised learning and deep learning, data algorithm based on the amount of training data have higher requirements, at the same time, SAR images are often limited to the amount of image data, therefore, this paper combines the SAR image imaging features of target azimuth sensitivity, puts forward two methods for image enhancement based on SAR data, through the object rotation the image in azimuth, and increase the random integer value in the original image to get more image data, the experimental results show that the proposed method is effective. (2) deep learning (Deep Learning DL) is a hot topic in machine learning, speech recognition in many fields such as image recognition, Natural Language Processing and other fields the breakthrough. Convolutional neural network (convolutional neural networks, CNN) is a kind of deep learning algorithm, based on CNN principle, training a depth CNN network model SARnet, SARnet contains two volumes Laminate, convolution kernel size is 7 x 7, two pool layer, two layer fully connected; after enhancement method combined with data set is 22 times of the original database of training, training of the network, the network model is trained by the feature extraction of the MSTAR database test set, the classification of feature extraction the SVM, the final recognition rate reached 95.68%, the recognition rate is higher than the other CNN model. (3) recently in machine learning, researchers have focused on learning some features from the unlabeled data, according to the SAR image feature extraction, the extracting regions of interest in the input image size 64 x 64; then using unsupervised K mean feature learning algorithm, combining the training database data after enhancement, learn some useful representation. By block self encoding and optimization parameters were accepted domain features of SAR image can make learning The model to learn the characteristics of variety, and proved by experiments, the mainstream recognition in unsupervised feature feature learning obtained can make SAR image recognition rate of 96.67% with the rate of K-means.

【学位授予单位】：五邑大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN957.52

【参考文献】