基于卷积神经网络的遥感图像配准方法

发布时间：2019-01-02 12:59

【摘要】：图像配准的主要目的是为了实现同一目标区域在不同时间、不同视角或不同传感器获得的图像数据在空间位置上重合,图像配准问题是地理信息学、影像医学、计算机视觉等众多应用领域中基础性问题。对于完成卫星遥感图像之间的配准,得出的配准信息对于完成目标识别、图像融合、场景重建等诸多应用问题的实现,有着至关重要的作用。在当前海量的遥感图像数据信息面前,传统的人工选取图像之间控制点实现遥感图像配准的方法已经无法满足实际应用中对于数据实时性的要求,所以改善自动化图像配准技术,已成为图像配准领域中的研究重点方向。传统的图像配准算法主要分为两大类:基于图像区域的配准算法和基于图像特征的配准算法。本文主要采用了基于局部特征的配准算法,并通过训练好的卷积神经网络来获取控制点的特征表达,以此来取得在图像配准的特征匹配阶段有较好的正确匹配对的数量,进而实现光学遥感图像之间的配准,本文验证了提出方法的可行性,本文主要完成的工作具体有下列几点:1.总结了图像配准技术现阶段的发展情况和传统的图像配准流程,并对未来图像配准技术的发展方向做出了展望;2.介绍了图像配准以及卷积神经网络的理论知识,并对卷积神经网络原理进行了详细的推导说明;3.采用最大稳定极值区域(Maximally Stable Extremal Regions,MSERs)提取训练卷积神经网络所需要的特征样本,并构造合适的网络结构进行网络的训练。4.利用训练完成的卷积神经网络模型转化待配准图像之间控制点的特征,并形成相应的特征表达,使用所得出的特征表达进行特征匹配。最后在光学遥感图像上进行了此方法的模拟仿真实验,并得出较好的图像配准效果。
[Abstract]:The main purpose of image registration is to realize the coincidence of image data obtained from the same target region at different time, different angle of view or different sensors in space. The problem of image registration is geographic informatics and image medicine. Basic problems in many fields of application, such as computer vision. For the registration of satellite remote sensing images, the registration information is very important to the realization of target recognition, image fusion, scene reconstruction and so on. In the face of the current massive remote sensing image data information, the traditional method of realizing remote sensing image registration by manually selecting control points between images can no longer meet the requirement of real-time data in practical applications. Therefore, improving the automatic image registration technology has become the focus of research in the field of image registration. Traditional image registration algorithms are mainly divided into two categories: image region-based registration algorithm and image feature-based registration algorithm. This paper mainly adopts the registration algorithm based on local features, and obtains the feature expression of the control points by the trained convolution neural network, so as to obtain the number of correct matching pairs in the feature matching stage of image registration. Then the registration of optical remote sensing images is realized. The feasibility of the proposed method is verified in this paper. The main work accomplished in this paper is as follows: 1. The current development of image registration technology and the traditional image registration process are summarized, and the future development direction of image registration technology is prospected. 2. The theoretical knowledge of image registration and convolution neural network is introduced, and the principle of convolution neural network is deduced in detail. The maximum stable extremum region (Maximally Stable Extremal Regions,MSERs) is used to extract the feature samples needed to train the convolutional neural network, and the appropriate network structure is constructed to train the network. 4. The convolution neural network model is used to transform the features of the control points between the images to be registered, and the corresponding feature expression is formed, and the obtained feature expression is used to match the features. Finally, the simulation experiment of this method is carried out on the optical remote sensing image, and a better image registration effect is obtained.
【学位授予单位】：南昌大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP751

【参考文献】