【作者】 滕鑫鹏；

【导师】 宋顺林；

【作者基本信息】 江苏大学 ， 计算机应用技术， 2014， 博士

【摘要】 高分辨率遥感图像的应用,可以使我们获取了更加精确、丰富和全面的信息。从遥感图像中抽取出信息,通过识别出感兴趣的目标获取知识,完成图像的理解是遥感图像应用的根本目标。遥感图像可以提供的信息中,道路信息是很重要的一部分。随着道路信息的不断更新,传统的人工操作已无法满足需求,于是,将遥感技术与电子技术及图像识别技术结合起来,研究遥感图像道路的自动提取,对于道路监控和GPS导航及地图及时更新都有重大意义,也是目前国内外研究的重点。论文主要研究遥感图像道路提取的方法,首先探讨了在线数据库中道路遥感图像的识别分类,然后对遥感图像云层进行自动识别的研究,以避免遥感图像中存在云层覆盖的情况,最后分别用两种方法对遥感图像中道路进行提取。针对在线数据库遥感图像分类,提出基于文本与图像信息融合的道路遥感图像识别分类的方法,需要先提取图像的文本特征与图像特征,再将二者的特征进行融合,通过支持向量机的训练,可得到较好的分类结果,该方法可移植到其他图像在线数据库分类识别与构建上。针对遥感图像云层,提取图像的纹理特征,选择4个纹理特征参数——角二阶矩、对比度、相关度和熵对云层图像进行自动识别。针对遥感图像道路,首先使用结构张量的方法进行主方向的计算,并改进了主方向的计算方法,结合Gibbs抽样对道路进行提取,该方法适用于有遮挡的道路,但无法说明道路的重要性；采用圆投影变换进行道路提取,通过与初始模板匹配找到最优模板来提取道路,该方法可以观察到道路的重要性。论文研究的具体内容如下：(1)基于文本与图像信息融合的道路遥感图像识别。在线图像数据库中的图像并不是在相同的实验室环境和相同的技术参数中统一测量的,因此,现有的方法并不能直接对道路遥感图像进行识别。针对这一问题,利用在线图像数据库中的道路遥感图像及其注释来获得高精度的道路遥感图像识别。利用空间金字塔关键字直方图来描述图像的特征,并通过融合图像和文本信息提高了道路遥感图像的识别精度。使用从图像数据库得到的图像信息和文本信息训练支持向量机以得到更高的分类精度,然后在整合所有信息后可以得到支持向量机的后验概率值和最终结果。相较于使用单独的图像特征或单独的文本特征进行识别,该方法具有较好的识别准确率和分类性能,可移植到其他图像在线数据库分类识别与构建上。(2)基于纹理特征对遥感图像云层自动识别。针对高分辨率遥感图像中云层的自动识别问题,提出一种基于图像纹理特征的云层自动识别方法,通过灰度共生矩阵来对图像中云层和下垫面的纹理特性进行统计分析,选择对云层和下垫面进行有效区分的4个纹理特征参数——角二阶矩、对比度、相关度和熵对图像进行识别,最后通过图像空间域的云层识别方法来对纹理识别结果进行修正,有效提高了云层识别的准确性,为遥感图像道路提取奠定基础。(3)基于结构张量进行遥感图像道路提取。针对基于结构张量的主方向计算方法计算结果不够精确的缺点,对高斯滤波进行改进并结合canny算子,提出了改进的主方向计算方法,然后基于局部主方向结合Gibbs抽样进行遥感图像道路提取。该算法适用于有遮挡的道路图像,可以比较精确地对道路进行提取。(4)基于圆投影变换进行遥感图像道路提取。本文中基于结构张量的遥感图像道路提取方法不能表明道路的关键性,采用圆投影变换理论进行遥感图像道路提取,可以比较精确的提取道路,同时从结果图中观察到道路的重要性信息。更多还原

【Abstract】 People can get more precise, rich and comprehensive information from high-resolution remote sensing image. The goal of the application of remote sensing image is getting information from remote sensing image, accessing knowledge by identifying the interested target, and at last understanding the image. With constantly update of the roads imformation, the traditional manual method can not meet demand, so the remote sensing technology, electronic technology and image recognition technology were combined to research the automatic extraction of the roads in remote sensing images. The research has great significance on road monitoring, GSP navigation and map update, and it also is the focus in the world.The paper researched road extraction methods of remote sensing image. Firstly, the classification of road remote sensing images in online database was researched; and then studied the automatic identification methods of remote sensing image of clouds in order to avoid clouds over the roads in the remote sensing image; at last, two methods were used to extract roads in remote sensing images. The classification method of road remote sensing images based on test and image information fusion was proposed to identify the road remote sensing image in the online remote sensing image database. The method extracted text features and image features, then made the features of both fusion, trained the features by SVM, and it would get better results by this method. Four texture feature parameters (angular second moment, contrast, correlation and entropy) are chosen to extract out to identify the remote sensing image of clouds. Structure tensor was used to calculate the main direction of the remote sensing image of road, and the improved calculation method was also proposed, and then connected with Gibbs sampling to extract the road in the image. This method was applicable to sheltered road, but can not explain the importance of the road. Then the circular projection transformation was used to extract roads, and this method can find the optimal template by matching the initial template to extract roads, it also can ixplain the importance of the road.The main research contents are as follows: (1) The road remote sensing image recognition based on fusing text and image information. The conventional method can not identify the road remote sensing image directly because the images in the online image database are not measured in the same laboratory environment and used with the same technical parameters. Road remote sensing images and their comments in the online image database can be used to get high precision road remote sensing image recognition results. Space pyramid keywords histogram was used to describe the features of the images, and the fusion of image information and text information can improve the recognition accuracy of the road remote sensing images. The information getted from online image database were trained by SVM to get better classification accuracy. At last, the final results and posterior probability values of SVM can be getted by integrating all the information. This method had better recognition accuracy and classification performance compared with using a separate image features or separate text features to identify the images.(2) Automatic recognition of remote sensing image of clouds based on texture features. The automatic identification method based on texture features was proposed for the high-resolution remote sensing image of clouds. The texture features of clouds and surface in images were statistical analyzed by gray level co-occurrence matrix; and then four texture feature parameters (angular second moment, contrast, correlation and entropy) of clouds and surface are chosen for the recognition of images; finally, the results was corrected by the clouds recognition method of image spatial domain.(3) Road extraction in remote sensing image based on structure tensor. The calculation method of main direction of road was improved by combined improved Gaussian filtering and Canny operator for the purpose of solving the results were not precise by the primitive method; and then extracted the road in the remote sensing image by combining local main directions and Gibbs sampling. This method can be more precise on the road extraction.(4) Road extraction in remote sensing image based on circular projection transformation. The method of road extraction in remote sensing image based on structure tensor can not explain the importance of the roads. Circular projection transformation can be used to extract roads in remote sensing images, and the method can get more precise result, at the same time, the importance of the roads can be ovserved from the result images.更多还原