基于自适应频域信息和深度学习的SAR图像分割
发布时间：2019-05-24 20:01
【摘要】：SAR图像包含多种地物信息,已经被广泛应用在目标识别、灾害评估、环境监测与跟踪等民用领域和军用领域,图像中各类目标的准确分割,对SAR图像后续的处理有重要的意义。SAR具有全时段、全天候的优势,但由于数据量庞大、图像地物目标丰富、受相干斑噪声污染严重等特点,高分辨SAR图像的处理面临着众多挑战。因此,SAR图像分割一直是SAR信息处理应用领域的热门研究内容之一。在研究目前SAR图像分割发展情况的基础上,本文考虑高分辨SAR图像的特点,结合自适应频域信息和深度学习理论进行SAR图像分割的研究,主要研究内容如下:1.简单分析了SAR图像的相干斑噪声和区域纹理特点,分析比较了现有的SAR图像分割的两大类方法的优缺点,根据两者的优缺点问题,提出了一种新的SAR图像分割方法,该方法以自适应窗口提取SAR图像的低高频信息作为特征向量,并对提取的信息进行改进的FCM分割,通过与几种传统方法结果的比较,证明了本方法的有效性。2.介绍了支持向量机在分类方面的优势,说明了图像分割在采用支持向量机算法的技术实现,采用凸壳性质的支持向量机方法进行图像分割,分析了非局部思想在图像处理方面的优势,用非局部思想提取频域信息,并将频域信息引入支持向量机的凸壳性质,提出了一种基于非局部频域信息和支持向量机的SAR图像分割。和经典方法对比实验结果表明,本方法能够达到更佳的分割效果。3.详细研究并介绍了深度学习CNN卷积神经网络,并用正交试验设计改进的PSO粒子群算法调整网络,提出了一种基于超像素和正交PSO修正深度学习的图像分割,和其他算法对比实验表明,本方法能够充分提取图像信息,并在一致区域和边缘上达到很好的分割效果。
[Abstract]:SAR images contain a variety of ground information, which have been widely used in civil and military fields such as target recognition, disaster assessment, environmental monitoring and tracking, and the accurate segmentation of all kinds of targets in images. Sar is of great significance to the subsequent processing of SAR images. Sar has the advantages of all-time and all-weather, but because of the huge amount of data, rich image objects, serious pollution by speckle noise and so on, Sar has the advantages of full time and all-weather. High resolution SAR image processing faces many challenges. Therefore, SAR image segmentation has always been one of the hot research contents in the field of SAR information processing. On the basis of studying the development of SAR image segmentation, this paper considers the characteristics of high resolution SAR image, combined with adaptive frequency domain information and depth learning theory to study SAR image segmentation, the main research contents are as follows: 1. The speckle noise and region texture characteristics of SAR images are briefly analyzed, and the advantages and disadvantages of the two existing SAR image segmentation methods are analyzed and compared. according to the advantages and disadvantages of the two methods, a new SAR image segmentation method is proposed. In this method, the low high frequency information of SAR image is extracted by adaptive window as feature vector, and the extracted information is segmented by improved FCM. Compared with the results of several traditional methods, the effectiveness of this method is proved. 2. This paper introduces the advantages of support vector machine in classification, and explains that image segmentation is realized by using support vector machine (SVM) algorithm, and image segmentation is carried out by using convex hull support vector machine (SVM). The advantages of nonlocal thought in image processing are analyzed. The frequency domain information is extracted by nonlocal idea, and the frequency domain information is introduced into the convex shell property of support vector machine. A SAR image segmentation based on non-local frequency domain information and support vector machine (SVM) is proposed. Compared with the classical method, the experimental results show that this method can achieve better segmentation effect. In this paper, the deep learning CNN convolution neural network is studied and introduced in detail, and an improved PSO particle swarm optimization algorithm is designed to adjust the network by orthogonal experiment. An image segmentation based on super pixel and orthogonal PSO modified depth learning is proposed. Compared with other algorithms, this method can fully extract image information and achieve good segmentation effect on consistent region and edge.
【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】：2014
【分类号】：TN957.52
本文编号：2485142
[Abstract]:SAR images contain a variety of ground information, which have been widely used in civil and military fields such as target recognition, disaster assessment, environmental monitoring and tracking, and the accurate segmentation of all kinds of targets in images. Sar is of great significance to the subsequent processing of SAR images. Sar has the advantages of all-time and all-weather, but because of the huge amount of data, rich image objects, serious pollution by speckle noise and so on, Sar has the advantages of full time and all-weather. High resolution SAR image processing faces many challenges. Therefore, SAR image segmentation has always been one of the hot research contents in the field of SAR information processing. On the basis of studying the development of SAR image segmentation, this paper considers the characteristics of high resolution SAR image, combined with adaptive frequency domain information and depth learning theory to study SAR image segmentation, the main research contents are as follows: 1. The speckle noise and region texture characteristics of SAR images are briefly analyzed, and the advantages and disadvantages of the two existing SAR image segmentation methods are analyzed and compared. according to the advantages and disadvantages of the two methods, a new SAR image segmentation method is proposed. In this method, the low high frequency information of SAR image is extracted by adaptive window as feature vector, and the extracted information is segmented by improved FCM. Compared with the results of several traditional methods, the effectiveness of this method is proved. 2. This paper introduces the advantages of support vector machine in classification, and explains that image segmentation is realized by using support vector machine (SVM) algorithm, and image segmentation is carried out by using convex hull support vector machine (SVM). The advantages of nonlocal thought in image processing are analyzed. The frequency domain information is extracted by nonlocal idea, and the frequency domain information is introduced into the convex shell property of support vector machine. A SAR image segmentation based on non-local frequency domain information and support vector machine (SVM) is proposed. Compared with the classical method, the experimental results show that this method can achieve better segmentation effect. In this paper, the deep learning CNN convolution neural network is studied and introduced in detail, and an improved PSO particle swarm optimization algorithm is designed to adjust the network by orthogonal experiment. An image segmentation based on super pixel and orthogonal PSO modified depth learning is proposed. Compared with other algorithms, this method can fully extract image information and achieve good segmentation effect on consistent region and edge.
【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】：2014
【分类号】：TN957.52
【参考文献】
相关期刊论文 前2条
1 邹焕新,蒋咏梅,匡纲要,郁文贤;一种基于斑点抑制的SAR图像舰船航迹检测算法[J];电子与信息学报;2003年08期
2 谭炳香,李增元;SAR数据在南方水稻分布图快速更新中的应用方法研究[J];国土资源遥感;2000年01期
，本文编号：2485142
本文链接：https://www.wllwen.com/kejilunwen/wltx/2485142.html
