Image Registration Procedure

Currently, it is difficult to find a universal method to cope with all registration situations, any registration algorithm must consider image imaging principles, geometric deformation, noise impact, registration accuracy, and other factors. However, in principle, the registration algorithm can be roughly divided into the following four steps:

(1) Feature Extraction

Manual or automatic methods are used to detect unchanged features in an image, such as closed areas, edges, outlines, and corner points. The Feature Extraction Algorithm must meet three conditions.

(A) significance. The extracted features should be obvious, widely distributed, and easy to extract;

(B) noise resistance, strong noise suppression capability and no sensitivity to changes in imaging conditions;

(C) Consistency, which can accurately detect the common features of the two images;

 

(2) Feature Matching

The relationship between extracted features is established through feature description counts and similarity measurements. The area gray scale, feature vector space distribution, and feature symbol description commonly used in feature matching. Some algorithms also complete the estimation of transformed model parameters while performing feature matching;

 

(3) transform model estimation

Based on the geometric distortion between the image to be registered and the reference image, the geometric transformation model that best fits the changes between the two images can be divided into a global ing model and a local ing model. The global ing model uses all control point information for global parameter estimation. The local ing model uses the local features of the image for local parameter estimation. Common transformation models include affine transformation, perspective transformation, and polynomial transformation, among which the most commonly used are affine transformation and polynomial transformation.

 

(4) Coordinate Transformation and Interpolation

Convert the input image to the corresponding parameter so that it is in the same coordinate system as the reference image. Because the coordinate points after image transformation are not necessarily integers, Some interpolation operations must be considered. Common interpolation methods include: nearest neighbor interpolation, bilinear interpolation, double cubic Interpolation, B-spline interpolation, and Gaussian interpolation;

Image Registration Procedure