基于无监督方法的SAR图像变化检测

发布时间：2018-01-07 09:08

本文关键词：基于无监督方法的SAR图像变化检测　出处：《西安电子科技大学》2014年硕士论文　论文类型：学位论文

【摘要】：合成孔径雷达(Synthetic Aperture Radar,SAR)技术具有高分辨率、全天时、全天候工作的特点,被广泛地应用于科学研究、工农业生产以及军事等领域。目前随着自然环境变化的加剧,城市也在迅速发展,各种自然灾害频频发生,图像变化检测技术也随之快速崛起。无监督方法对于解决大量的实际问题具有重要的意义。本文对基于无监督的SAR图像变化检测方法进行研究,对已有的多种方法进行改进,并提出自己的创新性算法,目前在该研究领域取得的研究成果如下:1.对无监督特征学习方法进行了研究。结合无监督特征学习方法以及SAR图像变化检测问题的特点,提出了一种全新的基于深度学习的SAR图像变化检测方法。该学习算法主要包括无监督的特征学习和有监督的调整过程,无监督的特征学习目的在于学习两幅图像之间的关系的表达,有监督的微调目的在于学习变化像素与未变化像素的概念。通过设计一个深度神经网络来检测变化与未变化的部分。主要思路为使用深度神经网络直接作用于两幅图像以获得最终变化检测结果,从而省去了生成差异图的过程,避免了差异图结果对最终变化检测结果的影响。算法在多组具有不同特点的数据集上进行实验,均可以得到满意的结果。2.对SAR图像变化检测问题进行分解研究。分别针对差异图的构造和差异图分类方法提出了改进的策略。该部分主要是基于改进的邻域比操作和修正的KI阈值算法实现的。改进的邻域比操作主要是通过结合灰度信息和邻域像素的空间信息产生差异图。为了进一步降低噪声的影响,根据邻域像素与中心像素的相似性关系,给每一个邻域像素赋一个合理的阈值。基于广义高斯假设的KI阈值算法主要是用来建模变化类与未变化类。实验结果表明,这两种改进方法的结合收到了良好的效果。3.对差异图的分析进行研究。使用多尺度的思想结合水平集分割算法解决SAR图像变化检测问题。通过定义能量函数来区分差异图的变化类与未变化类。根据水平集方法最小化能量函数寻找一个最优轮廓来分隔图像。为了获得对噪声更加鲁棒的初始化轮廓,使用平稳小波变换来产生多尺度的图像。在低尺度下,噪声和错误边缘将被减少,因此低尺度下的水平集演化对噪声更加敏感,且避免闭合轮廓陷入局部最优或错误边缘。低尺度上的结果将被作为更高尺度上水平集演化的初始化曲线。
[Abstract]:Synthetic Aperture radar (SAR) technology is widely used in scientific research due to its high resolution, all-weather and all-day work. Industrial and agricultural production as well as military and other fields. With the aggravation of natural environment changes, cities are also developing rapidly, a variety of natural disasters occur frequently. Image change detection technology is also rising rapidly. Unsupervised method is of great significance to solve a large number of practical problems. In this paper, the unsupervised SAR image change detection method is studied. The existing methods are improved and their own innovative algorithms are proposed. The research results in this field are as follows: 1. The unsupervised feature learning method is studied, which combines the unsupervised feature learning method and the characteristics of SAR image change detection problem. A new SAR image change detection method based on depth learning is proposed, which mainly includes unsupervised feature learning and supervised adjustment process. The purpose of unsupervised feature learning is to learn the expression of the relationship between two images. The purpose of supervised fine-tuning is to learn the concept of variable pixel and unchanged pixel. A depth neural network is designed to detect the changed and unchanged parts. The main idea is to use the depth neural network to act directly on the two parts. The image is used to obtain the final change detection result. Thus, the process of generating difference map is eliminated, and the influence of the result of difference map on the final change detection result is avoided. The algorithm is experimented on different sets of data sets with different characteristics. All of them can get satisfactory results. 2. Decomposing the problem of SAR image change detection. The improved strategy is put forward for the construction of difference map and the classification method of difference map respectively. This part is mainly based on the improved method. Neighborhood ratio operation and modified Ki threshold algorithm are implemented. The improved neighborhood ratio operation is mainly by combining gray information and neighborhood pixel spatial information to produce difference map. In order to further reduce the impact of noise. According to the similarity between neighborhood pixels and center pixels. Ki threshold algorithm based on generalized Gao Si hypothesis is mainly used to model change class and unchanged class. The experimental results show that. The combination of the two improved methods has a good effect. 3. The analysis of the difference map is studied. The multi-scale idea combined with the level set segmentation algorithm is used to solve the problem of SAR image change detection. The energy is defined. Based on the level set method to minimize the energy function to find an optimal contour to separate the image. In order to obtain more robust initialization contour for noise. Stationary wavelet transform is used to generate multi-scale image. At low scale, noise and error edge will be reduced, so the level set evolution at low scale is more sensitive to noise. The results on the lower scale will be regarded as the initialization curve of the evolution of the level set on a higher scale.
【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

【共引文献】