大幅面海洋卫星遥感图像目标检测研究

发布时间：2018-03-31 05:33

本文选题：高分辨遥感图像　切入点：背景统计建模　出处：《深圳大学》2017年硕士论文

【摘要】：海洋遥感技术是全球变化侦测和军事侦察等领域重要研究课题之一,如何有效精准地从大量的大幅面海洋遥感图像中提取出重要区域,是海洋遥感技术中的重要研究方向。其中,舰船目标检测是重点研究的内容,具有重要的军事和民用意义。但是由于遥感图像的空间分辨率越来越高,其尺寸越来越大,如何从大尺度大幅面遥感图像中进行海上目标检测存在一定的困难。所以本文针对大幅面海洋遥感图像目标检测中的一系列关键问题进行了深入的研究,包括系统的分析了大幅面海洋遥感图像的背景特性,提出了大范围的背景统计建模和模型参数的双层细化估计方法,以及提出了基于大范围背景模型的高分辨海洋遥感图像目标检测方法。论文的研究成果主要是以下三个方面:1)分析了卫星高分辨海洋遥感图像局部和大范围的海洋背景统计特性。利用局部和大范围特征统计量分析了卫星高分辨遥感图像中的局部和大范围背景统计特性,得出了海洋背景的统计均值与方差具有局部相似性和大范围连续变化性。2)建立了高分辨海洋遥感图像背景的大范围曲面高斯分布统计模型,提出了模型参数的双层细化估计方法。利用K-S假设检验对局部海洋遥感背景图像的灰度分布特性进行检验,得出了其局部海洋遥感背景图像灰度分布特性最符合高斯分布,结合海洋背景统计均值与方差的局部相似性和大范围连续变化性,采用曲面拟合的方法为海水背景空间中每个像素点建立统计分布模型,接着介绍了模型参数双层估计方法中的子图像参数粗估计和内部块参数细化插值,得到了大范围的遥感图像背景曲面高斯统计模型。3)设计了一种基于大范围背景模型的高分辨海洋遥感图像目标检测方法。首先将图像子块的均值和方差作为分类向量,利用模糊C均值聚类方法对纯海洋和非纯海洋的图像子块进行粗略分类预处理;接着采用该大范围的背景统计模型对大幅面海洋遥感图像中非纯海洋图像子块进行目标检测,得到大幅面遥感图像候选目标检测结果图;最后提取任意两个候选目标的质心距离和重叠率特征,利用模糊推理对候选舰船目标进行融合后处理,得到最终检测结果。
[Abstract]:Ocean remote sensing technology is one of the important research topics in the fields of global change detection and military reconnaissance. How to extract important areas from a large number of large format ocean remote sensing images effectively and accurately, It is an important research direction in ocean remote sensing technology. Among them, ship target detection is an important research content, which has important military and civil significance. However, because of the higher spatial resolution of remote sensing image, the size of ship target detection becomes larger and larger. There are some difficulties in the detection of large scale and large scale remote sensing images. Therefore, a series of key problems in large scale ocean remote sensing image detection are studied in this paper. The background characteristics of large format marine remote sensing images are systematically analyzed, and a large range of background statistical modeling and two-layer thinning estimation method for model parameters are proposed. And a high resolution ocean remote sensing image target detection method based on large range background model is proposed. The main research results of this paper are as follows: 1) analyze the local and large area sea of satellite high resolution ocean remote sensing image. Statistical characteristics of ocean background. Local and large-scale background statistical characteristics in satellite high-resolution remote sensing images are analyzed by using local and large-scale feature statistics, The statistical mean and variance of ocean background have local similarity and wide range continuous variation. 2) A statistical model of Gao Si distribution on large scale curved surface with high resolution ocean remote sensing image background is established. A two-layer thinning estimation method for model parameters is proposed. The gray distribution characteristics of local marine remote sensing background images are tested by K-S hypothesis test, and the results show that the gray distribution characteristics of local marine remote sensing background images are most consistent with Gao Si distribution. Combined with the local similarity between the statistical mean and variance of ocean background and the continuous variation in large range, a statistical distribution model for each pixel in sea water background space is established by using curved surface fitting method. Then, the rough estimation of sub-image parameters and the interpolation of internal block parameters in the two-level estimation of model parameters are introduced. In this paper, Gao Si's statistical model of large-scale remote sensing image background surface. 3) A high resolution ocean remote sensing image target detection method based on large-scale background model is designed. Firstly, the mean and variance of image subblocks are taken as classification vectors. The fuzzy C-means clustering method is used to preprocess the image subblocks of pure and non-pure ocean roughly, and then the large scale background statistical model is used to detect the sub-blocks of non-pure ocean images of large scale ocean remote sensing images. Finally, the centroid distance and overlap rate of any two candidate targets are extracted, and the final detection results are obtained by using fuzzy inference to fuse the candidate ship targets.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】