Research On Image Registration Based On Multi-scale Feature Points Extraction

Image registration is a basic problem of image processing, which is the prerequisite and key step to do further image analysis. Image registration refers to the techniques of image processing that capture two or more images in the same scene by the different sensors at different time or from different viewpoints. After years of research, image registration has been widely applied in image fusion, image tracking and remote sensing fields. Recently image registration algorithms can be divided as follows:based on image pixel values, based on transform and based on image features. It is showed by experiences that image registration algorithm based on feature extraction only analyse image features instead of analysis of the whole image piexels, thus greatly reduced the computation of image processing. Now, it is a main stream of doing image registration research. Therefore, this paper focuses on the study of the image registration method of feature extraction--feature points extraction.In this paper, firstly studied and anaylzed some classic feature point extraction algorithms:Moravec operator, SUSAN operator and Harris operator, then verified and compared those three operators’advantages and disadvantages. Secondly, as a reason of Harris operator can extract image point features more uniform, reasonable but does doesn’t have the performance of scale invariance, the paper introduced the idea of multi-scale feature points detection. The paper focused on the research of multi-scale feature point detection algorithm in scale space:Harris-Laplace algorithm and the SIFT descriptor. For the Harris-Laplace algorithm which brought many redundancy points of the same local feature in multi-scale space, the paper presentd the improved Harris-Laplace algorithm which based on the idea:optimal local structure represented and redundancy points excluded. Thirdly, the paper eliminated the coarse error matched feature point pairs, estimated the transformation matrix through random sample consensus (RANSAC) algorithm and implemented the image registration alogrithm. Finally, the paper realized the basic image fusionuses using improved image registration algorithm which the paper proposed and achieved good resultsImage registration is mainly to improve the algorithm processing speed and accuracy, but due to different applications and scenarios has different requirements, so a common algorithm is not yet solved. So the next step is easy and accuracy to realize the image registration. It is showed by experiments that the improved Harris-Laplace algorithm performed higher accuracy, faster, stronger robustness than SIFT algorithm. The experiments showed that the improved Harris-Laplace algorithm proposed by the paper has the following advantages compared to the classical Harris algorithm and Harris-Laplace algorithm:keep the advantages of classic Harris that feature point distribution reasonable and overcomed its shortcomings of scale sensitive. The imporved alogrithm owned higher position accuracy, faster match speed and more robust compared to the original Harris-Laplace algorithm.