Image Registration Overview

Image Registration is widely used in target detection, model reconstruction, motion estimation, feature matching, tumor detection, lesion location, angiogram, geological exploration, aerial reconnaissance, and other fields.

Each registration method is usually designed for a specific problem. Among the many methods, the only commonality is that each registration problem will eventually find the most transformation in the transformation space, this transformation can make the two images match in a certain sense. However, for different application fields, different requirements on image types require specific analysis of specific problems.

Based on the relationship between the images to be registered, some researchers divided the image registration into four categories: multi-source image registration, template-based registration, multi-angle image registration, and Time Series Image Registration. For details, see

 

At present, it is commonly used to classify the registration image information according to the processing methods in the image registration algorithm, which can be divided into three types:

(1) registration method based on gray information of the image to be registered

Based on the gray information of the entire image, a similarity measurement between the image to be registered and the reference image is established. A search algorithm is used to find the transformation model parameters that make the similarity measurement reach the optimal value. It is also called direct registration;

 

(2) registration method based on the domain information of the image to be registered

Generally, the domain name registration is performed based on Fourier transform. Fourier transform can be used in image registration with translation, rotation, and scaling because

(A) After Fourier transformation, the Phase Relationships of the two images with translation volumes are different because the translation volume in the time domain directly determines the Phase Relationships in the Fourier transformation domain;

(B) For two images with a rotation volume, the rotation volume in the Fourier transform area remains unchanged;

(C) For two images with scale scaling, first convert the coordinate system to the logarithm coordinate system. The image scale can be converted to image translation for processing.

 

(3) Registration Method Based on Image Feature Information to be registered

Feature-based registration is currently one of the most common registration methods. This algorithm only needs to extract the vertices, lines, edges, and other feature information in the image to be registered without other auxiliary information, while reducing the amount of computing and improving the efficiency, it is able to have a certain degree of robustness to the changes in the gray scale of the image. However, because the algorithm only uses a small part of the feature information of the image, this algorithm requires a high accuracy and accuracy for Feature Extraction and feature matching, and is very sensitive to errors.

Based on the selected feature information, the feature-based image registration method is divided into three types

 

(A) Matching Based on Feature Points

Generally, the selected feature points are pixels with some singularity relative to their fields. Feature Points are often easy to be extracted, but feature points contain less information and can only reflect their location and coordinate information in the image, therefore, finding matching feature points in two images is the key.

 

(B) feature region-based matching

Find some obvious regional information in the image as the feature area. However, after finding the feature area in practical application, the most important thing is the regional heart point, therefore, this type of algorithm requires that the Feature Area Extraction accuracy be very high.

 

(C) feature edge-based matching

Edge is the most obvious feature in the image, and edge features are also one of the best features to be extracted. Therefore, edge matching methods are highly robust and suitable for a wide range of applications. However, these methods have high requirements for edge feature extraction and require edge information to be expressed in mathematical languages.

Image Registration Overview