Image Registration

Feature Detection and Extraction

For image registration, a sufficient number of control points (common features) are required to estimate an optimal geometric transformation between two images. The control points can be [Page 1539]selected manually or extracted automatically. They are, normally, any of the following features:

Figure 1 Image registration for mosaicking through overlaying the control points

Source: Authors.

• Line intersections
• Points of locally maximum curvature (such as building corners)
• Gravity centers of closed boundary regions (such as centers of building roofs or traffic islands)
• Centers of windows having locally maximum variance

The minimum number of control points required depends on the transformation function used. For example, when the conformal transformation function (Equations 1 and 2) is used, at least two control points are required. Because there are four unknowns a, b, c, and d in the transformation function, two control points are the minimum number to solve the equations and find the unknowns. More control points are usually required to achieve accurate solutions by least square approximation.

Feature Matching

After automated feature detection, numerous features can be extracted. However, many feature points may be extracted from one image but not from the other. Therefore, feature matching is necessary to find corresponding points (common features) in both images. The feature matching usually starts from one feature point on one image and then searches for the corresponding point on the other image.

Basically, there are two kinds of feature matching approaches: (1) area-based methods and (2) feature-based methods. In area-based methods, cross-correlation is often used as a similarity measure to find the corresponding point on the other image. In feature-based methods, the sum of squared differences is usually used to identify the corresponding point.

Transformation Function Selection and Fitting

Transformation function is used to model the geometric relationship between two images. A transformation function should take the possible geometric distortion between two images into consideration. There are two categories of possible

...