基于LDA的SIFT算法在遥感图像配准中的研究与应用

发布时间：2018-06-20 08:11

本文选题：图像配准 + SIFT　；参考：《江苏科技大学》2014年硕士论文

【摘要】：遥感是获得全球实时动态数据的一种最重要的途径。随着社会经济技术的发展，人们对实时动态数据的精确性和高效性有了更高层次的要求，以往的图像配准技术已经远远不能满足人们的需求，因此遥感图像的配准技术亟待提高，从而成为高效化、智能化、快速化的图像配准技术。传统的SIFT（Scale-invariant Feature Transform）算法是一种基于特征的配准方法。SIFT特征匹配算法最初是由David G. Lowe于2004年在总结了现有的基于不变量技术的特征检测方法的基础上提出的。SIFT算法具有诸多优点，能提取稳定的特征，也可以处理两幅图像之间发生旋转、平移、视角变换、仿射变换、光照变换等情况下的匹配问题，甚至在某种程度上对任意角度拍摄的图像也具备较为稳定的特征匹配能力，从而能够实现两幅差异较大的图像之间的特征的匹配。因此这种算法很适合在图像配准中使用。虽然SIFT算法具有精确度高、匹配能力强等许多优点，但是SIFT算法本身复杂度高，匹配时间久，尤其对于数据量较大的遥感影像的处理，应用SIFT算法时处理速度会明显降低，过于耗时。因此，将SIFT算法直接应用于遥感影像的处理并不实用，，在实际生活中也未能使用这种算法处理遥感影像。但是SIFT算法具有其他图像配准算法难以媲美的优点，有着非常好的应用前景。为此，我们通过大量研究提出了一种基于LDA算法的改进方案。即在SIFT算法的特征提取中加入线性鉴别分析方法（Linear DiscriminantAnalysis，LDA），以减少SIFT特征提取的维度。本文对提出的基于LDA的SIFT算法的特征原理和配准步骤进行了重点叙述。首先利用SIFT算法提取出图像的特征点向量，然后用线性鉴别分析方法对其进行特征抽取并降维。通过对遥感图像、高维自然图像和单幅人脸图像进行实验，实验结果表明LDA-SIFT算法在保证匹配精度的同时，实时性要优于传统的SIFT算法，其匹配时间相对于传统SIFT算法明显缩短。此方案既能保持SIFT算法本身的精确度高、匹配能力强的特点，又能提高匹配的效率，实时性较强。非常适合应用于维数高的图像配准，尤其是对时间和精度要求较高的遥感图像的配准。论文的主要研究成果如下： (1)论文分析了SIFT算法的特征原理以及线性鉴别分析方法LDA的理论基础。并且针对本文提出的改进算法具体原理以及图像配准的步骤进行描述。利用LDA易于实现的特点，将其和SIFT算法进行结合，从而实现全局优化，提高算法的可行性。 (2)将改进的算法通过编程设计，并将其应用于遥感图像配准、自然图像配准和人脸图像配准，通过实验证明，本算法针对数据量较大的遥感图像进行匹配时，能够解决传统算法实时性差的问题；针对自然、人脸图像进行匹配时，能够在提高匹配速率的同时，提高匹配算法的鲁棒性。本文提出的改进的SIFT算法，即LDA-SIFT算法既能降低遥感图像的配准时间，又极大的提高了配准效率，更好的满足了当今社会人们对实时动态数据的精确性和高效性的要求。
[Abstract]:Remote sensing is the most important way to obtain the global real-time dynamic data. With the development of social and economic technology, people have a higher level of requirements for the accuracy and efficiency of real-time dynamic data. The previous image registration technology has been far from meeting the needs of people. Therefore, the registration technology of remote sensing image needs to be improved, so that the registration technology of remote sensing image needs to be improved. It has become an efficient, intelligent and fast image registration technology.
The traditional SIFT (Scale-invariant Feature Transform) algorithm is a feature based registration method based on the.SIFT feature matching algorithm originally proposed by David G. Lowe on the basis of the existing feature detection methods based on the invariant technology in 2004. The algorithm has many advantages, which can extract stable features and can also be used. In the two images, the matching problem of rotation, translation, angle transformation, affine transformation, illumination transformation and so on, and even to some extent, has a more stable feature matching ability for the images taken at any angle, which can achieve the matching of features between two different images. Although the SIFT algorithm has many advantages, such as high accuracy and strong matching ability, the SIFT algorithm itself has a high complexity and long matching time, especially for the processing of remote sensing images with large amount of data, and the processing speed of the application of SIFT algorithm will be significantly reduced and time-consuming. Therefore, the SIFT algorithm is applied directly to remote sensing. Image processing is not practical, and this algorithm has not been used in real life to deal with remote sensing images. However, the SIFT algorithm has the advantages of other image registration algorithms, and it has a very good application prospect.
To this end, we put forward an improved scheme based on LDA algorithm through a lot of research. That is, adding the linear discriminant analysis (Linear DiscriminantAnalysis, LDA) in the feature extraction of the SIFT algorithm to reduce the dimension of the feature extraction of SIFT. This paper focuses on the feature principle and registration steps of the SIFT algorithm based on LDA. Firstly, the feature point vector of the image is extracted by SIFT algorithm, and then the feature extraction and dimension reduction are carried out by the linear discriminant analysis method. The experimental results show that the LDA-SIFT algorithm is better than the traditional SIFT calculation by using the remote sensing image, the high dimensional natural image and the single face image. The matching time is significantly shorter than that of the traditional SIFT algorithm. This scheme can not only maintain the high accuracy of the SIFT algorithm, but also improve the matching efficiency and the real-time performance. It is very suitable for high dimension image registration, especially for the registration of remote sensing images with higher time and precision.
The main research results of this paper are as follows:
(1) the paper analyzes the principle of the SIFT algorithm and the theoretical basis of the linear discriminant analysis method LDA, and describes the concrete principle of the improved algorithm and the steps of the image registration proposed in this paper. Using the features of the easy realization of LDA, the algorithm is combined with the SIFT algorithm, thus the global optimization is realized and the feasibility of the algorithm is improved.
(2) the improved algorithm is designed by programming and applied to remote sensing image registration, natural image registration and face image registration. Through experiments, it is proved that this algorithm can solve the problem of the poor real-time performance of the traditional algorithm when matching the large data of remote sensing images. At the same time, high matching rate improves the robustness of the matching algorithm.
The improved SIFT algorithm proposed in this paper, that is, LDA-SIFT algorithm can not only reduce the registration time of remote sensing images, but also greatly improves the registration efficiency, which can better meet the demands of people in today's society on the accuracy and efficiency of real-time dynamic data.
【学位授予单位】：江苏科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【相似文献】