An Automatic Registration For Multispectral Remote Sensing Images With Improved SIFT

Image registration is a classical problem to get the geometric correction encountered inmany image processing application such as sensing, medical, computer vision and so on, whereit is necessary to perform joint analysis of two or more images of the same scene acquired bydifferent sensors, or images taken by the same sensor but at different times with differentconditions. In the field of remote sensing, image registration is the crucial step in realizing imagefusion, area detection, scene rebuilding, and object detection and so on.The traditional procedure of manually registering satellite images requires to manualselection of control points in each image, which can’t meet real-time and automatic imageprocess system when a large number of images need to be registered. Thus, there is a need forautomated techniques that require little or no operator supervision. The current automaticregistration algorithms can be divided into area-based method and feature-based method. Theautomatic methods based on local feature are the focus in this paper, and we have completed thefollowing works:1. Introduce the image registration concepts, application, direction and the common steps ofregistration method.2. Introduce a series of extraction operators of feature points, and study the characteristic ofHarris operator, Affine-Harris operator, MSERs and Salient Region by analyzing theirperformance and applicability.3. Aiming at the need for high precision, introduce the step of SIFT operator, includingbuilding the DOG scale space, determination of keypoint location and keypoint descriptorassignment. But the standard SIFT is the lack of control on the spatial and scale-space distribution of the extracted feature impact the scale-invariant of feature and the precision ofglobal geometric transformation, and the feature is unstable due to the manual threshold Tr. Forthose problems, through pre-defined the number of features and statistics the information entropyof image of local instead of manual threshold, then compared the RMSE and running timebetween the standard SIFT and improved SIFT, the results shows that the improved SIFT makemore precision in geometric correction and less time in portion images.