Research And Application Of Automatic And Fast Multi-band Remote Sensing Image Registration Method Based On Improved SIFT

Due to the influence of the characteristics of remote sensing technology and sensor, there is somegeometric distortion in the initial remote sensing image. Image registration is one of the necessarypreprocesses before the following process of the remote sensing images. Image registration methods aredivided into three categories: gray information-based approach, transform domain-based and feature-basedapproach. Scale Invariant Feature Transform (SIFT) which used in this paper is one of the feature-basedapproaches. Image features descriptor generated by the SIFT algorithm keeps invariance to illuminationchanges, rotation and scale scaling, maintains a certain degree of stability to affine transformation,perspective changes and noise, etc. It’s suitable for image feature extraction. However, when we use theSIFT algorithm for feature points extraction of remote sensing image registration, because of thecharacteristics of the remote sensing images’ multi-band and large amount of data, it’s instability to extractremote sensing image feature points with image combined by different image bands and time-consuming ofthe whole calculate of the algorithm. As a result, it’s certain difficult to large-scale expand and application.This article will focus on these two problems of research and propose improvement method.The work done by this thesis mainly contains the following:Firstly, a feature point extraction method based on integrated application of multi-band remote sensingimage information was proposed. According by a large number of tests, we found that there is a greaterdifference on the number and position of the feature point, when using SIFT to extract feature point by thecombined image of different bands of remote sensing images. Therefore, I have extracted feature points offour typical band combinations firstly, and then designed a combined and remove repeat method to combine the final set of feature points, to enhance the number and quality of the feature point extraction.Secondly, an improved method of the build of Gaussian Scale-space was presented. I have analysed indepth of the image convolution method used by the constructed of Gaussian Scale-space in SIFT algorithm,and found that there are a lot of repeated multiplication operation. Considering the abundant of thecomputer hardware, we use more memory space to save calculate time. This paper calculate the weightingimages of the original image firstly, and then implement the convolution operation by the sum of pixel grayvalue which read directly in the weighting images. The amount of calculation of the multiplicationoperation in the improvement method is less than a quarter of the original, which can improve the GaussianScale-space construction of speed significantly.Thirdly, accelerate the improved SIFT algorithm in parallel by CUDA technology. On the basis ofin-depth study of the SIFT algorithm, I used the CUDA technology to accelerate the improved SIFTalgorithm, which put the time-consuming calculation of image processing operations on the GPU. It canenhance the execution efficiency of SIFT algorithm.Finally, some experiments were designed to verify forward improved algorithms. In this paper, I haveprepared a large number of different size, different regions of experimental data, designed some test schemefor the improved SIFT algorithm, and formulated appropriate evaluation criteria. The result of the test ofthe experimental zone data show that, improved algorithm enhances the effect of the extraction of featurepoints’ quantity and quality obviously, and reduces the time-consuming of the extraction of feature pointsof large-size remote sensing image substantially. At last, it’s good for remote sensing image registration inthe practical application of the improved SIFT algorithm.