Application Of LDA-based SIFT Algorithm In Remote Sensing Image Registration

Remote sensing is one of the most important approaches to obtain real-time dynamicdata around the world. With the development of social economy and technology，theefficiency and accuracy of the real-time dynamic data have a higher level of requirements.Previous image registration technology already cannot satisfy the demand of people. Soremote sensing image registration techniques need to be perfect, and finally become a kindof efficient, intelligent, and quick registration.The traditional SIFT（Scale-invariant Feature Transform)is a registration method basedon feature. SIFT feature matching algorithm is proposed by David G. Lowe in2004, it israised after summarizing the existing feature invariant methods. SIFT algorithm can extractthe characteristics of stability, also can deal with the matching problem between two images.For example, rotation, translation, perspective transformation, affine transformation,illumination change, etc. Even in any angle images to some extent also have a more stablefeature matching ability, so that they can achieve two different images between thecharacteristics of the match. Therefore, the proposed work is very suitable for imageregistration. It has the characteristics of high precision, strong matching ability. Althoughthis algorithm has many advantages, the algorithm itself is complex and needs a long time tomatch, especially for large amount of data of remote sensing images. When using SIFTalgorithm can significantly reduce processing speed. Therefore, it is not practical to processremote sensing images by traditional SIFT. But the SIFT algorithm is better than otherimage registration algorithm, so it has broad prospects.LDA-based SIFT algorithm is proposed. We combine LDA with SIFT in order toreduce the dimension of SIFT feature extraction. First, the original matching vectors ofimages are obtained by SIFT. Second, the process of dimension reduction and featureextraction of matching vectors is completed by LDA. The matching experiments to highdimensional remote sensing images, natural images and single face image are implementedwith the new algorithm, and the results show that the new algorithm can enhance thematching speed effectively as the matching precision is ensured, it saved a lot of time. Thisscheme can maintain the advantages of high precision, strong matching ability, also canimprove the matching efficiency. It is very suitable for high dimension of image, registration,especially for remote sensing images with high accuracy requirements. This dissertation made the following research results:(1)The dissertation analyzed the principle of SIFT algorithm and the theory ofLDA.The principles and methods about remote sensing image registration are introduced inthis dissertation.We combine LDA with SIFT for the reason that LDA is easy toimplement.So the global algorithm is optimized.(2)The improved algorithm is carried out by programming. The performance is verifiedby matching experiments to high dimensional remote sensing images, natural images andface images.The results show that the new algorithm can enhance the matching speedeffectively as the matching precision is ensured, it also improved the robustness of thematching algorithm.The improved SIFT algorithm is proposed in this dissertation, not only can reduce theremote sensing image registration time, but also greatly improved the efficiency ofregistration, well to meet the accuracy and efficiency of real-time dynamic datarequirements.