目标识别的遥感图像超分辨率方法研究

发布时间：2018-05-06 04:23

本文选题：图像配准 + 超分辨率重建　；参考：《哈尔滨师范大学》2014年硕士论文

【摘要】：1972年，美国发射了首颗地球观测卫星—landsat-1，标志着人类进入了遥感时代的新纪元。随着遥感技术的迅猛发展和广泛应用，虽然遥感图像的分辨率已经不断提高但还是难以满足许多领域对于高分辨率遥感影像的需求，比如在遥感图像小目标（飞机、船只、桥梁等）识别当中，需要寻找的目标通常都是米级目标，利用现有的遥感图像，进行小目标分割精度偏低。提高遥感图像的分辨率可以通过改良硬件设备实现，但工艺水平越复杂精细，成本也会越高昂，并且成像设备改良的空间也是有极限的，人们开始尝试利用软件的方法解决这一问题，超分辨率重建技术在这个背景下应运而生。本文利用超分辨率重建技术提高现有遥感图像的分辨率，提出了一种基于的超分辨率重建方法，为目标分割提供包含更丰富有效特征的遥感图像源数据，将超分辨率重建技术运用于小目标识别，并通过实验成功提出小目标（塘坝）。本文对超分辨率重建技术关键技术做了较为深入的研究，主要包括以下几个方面：遥感图像配准，高精度的图像配准是完成超分辨率重建的基础，本文介绍了现在常用的图像配准技术，并分析了现有遥感配准技术的难点问题，针对小目标遥感图像的特殊性采用一种基于SIFT的多光谱遥感图像配准方法，通过实验验证本文方法可以快速、自动的完成高精度配准。遥感图像超分辨率重构，目前序列超分辨率算法主要分为频域类和空域类两类方法。本文概述了两类方法的主要算法原理和优点，并提出一种基于Hopfield神经网络的超分辨重建算法，在遥感图像超分辨率重建的结果中，往往图像的边界等细节容易产生模糊，而这些部分又包含较多重要的信息，该方法可以有效提高细节保护的能力，通过实验取得较好的超分辨效果。将超分辨率技术运用在小目标分割中，小目标分割的精度不仅依赖于目标分割方法，高分辨率的遥感图像作为源数据也是重要因素之一，高分辨率的遥感图像能够为目标识别提供更多有效特征，提高识别精度。本文将超分辨率技术运用于目标分割中，，通过实验完了对小目标（塘坝）的目标识别。
[Abstract]:In 1972, the United States launched its first Earth observation satellite-Landsat-1, marking a new era in the era of remote sensing. With the rapid development and wide application of remote sensing technology, although the resolution of remote sensing images has been continuously improved, it is still difficult to meet the needs of high-resolution remote sensing images in many fields, such as small targets (aircraft, ships) in remote sensing images. In the recognition of bridges, the targets that need to be looked for are usually meter targets. The segmentation accuracy of small targets is low by using the existing remote sensing images. Improving the resolution of remote sensing images can be achieved through improved hardware, but the more sophisticated the process, the higher the cost, and there is a limit to the space for improved imaging equipment. People began to use software to solve this problem, and super-resolution reconstruction technology came into being under this background. In this paper, the super-resolution reconstruction technique is used to improve the resolution of the existing remote sensing images, and a super-resolution reconstruction method based on super-resolution is proposed, which provides the source data of remote sensing images with more effective features for target segmentation. The super-resolution reconstruction technique is applied to small target recognition, and the small target (Tangba) is proposed successfully by experiment. In this paper, the key technologies of super-resolution reconstruction are studied, including the following aspects: Remote sensing image registration and high-precision image registration are the basis of super-resolution reconstruction. This paper introduces the common image registration technology and analyzes the difficult problems of the existing remote sensing registration technology. According to the particularity of small target remote sensing image, a multi-spectral remote sensing image registration method based on SIFT is adopted. The experimental results show that this method can achieve high precision registration quickly and automatically. In remote sensing image super-resolution reconstruction, the sequence super-resolution algorithms are mainly divided into two categories: frequency domain and spatial domain. In this paper, the principle and advantages of the two kinds of methods are summarized, and a super-resolution reconstruction algorithm based on Hopfield neural network is proposed. In the super-resolution reconstruction of remote sensing image, the details such as the edge of the image are often blurred. These parts contain more important information. This method can effectively improve the ability of detail protection and obtain better super-resolution effect through experiments. The super-resolution technology is applied to small target segmentation. The accuracy of small target segmentation depends not only on the method of target segmentation, but also on the high resolution remote sensing image as one of the important factors. High resolution remote sensing images can provide more effective features for target recognition and improve recognition accuracy. In this paper, the super-resolution technique is applied to target segmentation, and the target recognition of small target (Tangba) is completed through experiments.
【学位授予单位】：哈尔滨师范大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】