Study On Feature-Based Image Registration Method

Image registration is a fundamental task in image processing used to (geometrically) match two or more images, for example, at different times, from different sensors or from different view points. Image registration is the foundation of many image processing and applications such as object recognition, change detection, 3-modeling, and the effect of registration directly works on the following processing steps.Now, the proposed approaches of image registration are either based on grey correlation or based on features. Gray-based approaches have some disadvantages, such as complexity of computation, sensitivity to gray, rotation, distortion and covering, and so on, but feature-based approaches can overcome these disadvantages, so they applied widely to image registration.But, there are mainly two difficulties consisting in feature-based approach: variety of detected features and complexity of computation of similiarity. To solve these two difficulties, this paper is made of several components following: First, comparing the approaches of edge detected, wavlet multi-scale product is used to detect edge and feature points of the image, at the same time, angles of the feature points are oberved.Second, based on edge image and feature points detecet by using wavelet multi-scale product, combining the virtues of alignment metric, a new method for image registration is proposed based on alignment metric of edge feature points. The first step is to detected edge image and feature points by using wavelet multi-scale product. In this step, wavelet multi-scale product can ensure the accuracy of edge image and feature points. The second step is that Alignment Metric of feature points can be calculated based on rotating angle getting from Angle Histogram. Then mapping points can be confirmed accurately by an appropriate threshold. The calculation of alignment metric only need to be based on edge images, so its complexity is very low.Third, combining image edge with mutual information, a new method for image registration is proposed based on mutual information of feature points. Only the mutual informations of pairs of feature points are need to compute, so its computation is very low. And another virtue of this mothed is we needn't search the most mapping pair of feature points, so local extremun is avoided.