Machine Vision: Classification of Image Registration Methods

Compared with the original image, the image to be registered has spatial transformation relationships such as offset, rotation, and scale. Image registration is the process of matching multi-spectral, multi-band images of the same scene acquired by different sensors or two or more images of the same scene acquired by the same sensor at different phases, orientations, and conditions (climate, illumination, camera position and angle, etc.).

1. Overview of Image Registration

1. Commonly used image transformations

1. Rigid body transformation

The distance between two points in one image remains unchanged after being transformed into another image.

2. Affine Transformation

A straight line on one image remains a straight line when mapped to another image after transformation, and they remain parallel.

3. Projection Transformation

A straight line on one image remains a straight line when mapped to another image after transformation, but the parallel relationship is basically not maintained.

4. Nonlinear transformation

A straight line on one image may not be a straight line when mapped to another image, but may be a curve.

2. Classification of image registration methods

1. Pixel-based registration method

Generally, the registration parameters are calculated based on the correlation function of the registration images, Fourier transform and other relationship expressions. The simplest method is the window (template) matching method.

2. Feature-based registration method

It mainly uses various algorithms to extract the features of the reference image and the real-time image, such as edges, corners, curvatures, invariant moments, etc., and then describes the extracted features so that they can be matched according to some similarity measure.

3. Model-based registration method

The image is registered in a nonlinear correction manner based on the mathematical model of image distortion, which is mostly used in medical images.

2. Image Registration Method

1. Registration and positioning method based on feature points

The number of matching points of the image pair gradually decreases as the threshold increases, but the matching points are more stable. SIFT features are local features of images with rich information. They are suitable for fast and accurate matching in massive feature databases, and reliable and stable feature matching points can be found by selecting appropriate thresholds.

2. Registration and positioning method based on image moments

The reference image and the image to be registered are preprocessed with feature extraction, the centroid principal axis method is used to complete the matching between the features of the two images, the registration mapping relationship between the images is established, and the image to be registered is transformed by affine transformation to achieve image registration and positioning.

The reference image and the image to be registered are processed separately, and the continuous and complete contours of the two images are obtained through threshold segmentation, median filtering, contour extraction, and contour tracking; the principal axes of the target contours of the two contour images are calculated respectively, and the angle between the two principal axes is the rotation angle of the image.