Image registration is one of the basic tasks of image processing, as early as the 70 's, people began to study aspects of image registration, from the most simple template matching correction image translation, to 90 mid-decade start for multi-modal image Quasi-wide research. In recent years, the research on registration technology has covered many application fields, and registration technology plays an important role in computer vision and pattern recognition, medical image analysis, remote sensing data processing, robotics, computer-aided design and manufacturing, astronomy and other disciplines. In the first three application fields, the research on image registration technology has been expanded. Image registration has become an essential part of many research topics, and it has become a bottleneck for improving accuracy and effectiveness in various problems.
Image registration is the process of matching and superimposing two or more images of the same scene acquired at different times, different sensors ( imaging equipment ) or under different conditions ( weather, illuminance, camera position and angle, etc. ) . A basic problem in the processing field. At present, image registration technology has been widely used in many fields, especially in several application fields of image fusion, image tracking and satellite remote sensing system.
Image registration technology has played a very important role in many fields, so the research on image registration technology is also extremely important. After years of research on image registration technology , certain research results have been achieved. The current image registration technology can be divided into manual registration and automatic registration. The manual registration method has been widely used in practice, but a large number of position control points need to be selected. This is a very boring, labor-intensive, repetitive and time-consuming work, and manual registration requires the operator’s experience. Human subjectivity makes it difficult to guarantee the accuracy of image registration. When high-precision and accurate image registration and image analysis are within a limited time, it is necessary to find an automated technology with little or no manual operation to register multi-modal images. Automatic image registration is a registration technology that does not require human intervention during the entire registration process, and it is also the ultimate development goal of image registration technology.
1.2 Domestic and foreign development status and trends of related technologies
Image registration is a very important research direction in the research of image processing. In the process of machine recognition, it is often necessary to align two or more images of the same scene acquired by different sensors at different times and under different imaging conditions in space, or to another image according to a known pattern Find the corresponding mode in the image registration, which requires image registration. Image registration is to compare and match the template with the image to be detected, and give a calculation result describing the degree of matching. If the calculation result of the algorithm shows that a certain part of the image is the same as or similar to the template and is greater than the set threshold, the matching is considered successful.
Early image registration technology is mainly used for the registration of multi-band remote sensing images after geometric correction, which is achieved by finding the extremum of the cross-correlation function. For example, in remote sensing image processing, the multi-spectral images of the same scene with sensors of different bands are registered correspondingly according to the nature of the image point, and then the features are classified according to the nature of the image point. If two images taken on the same ground at different times are used After registering the photos, find out the points where the characteristics have changed, and then you can analyze which parts of the picture have changed. The research of image registration involves many related knowledge fields, such as image preprocessing, image sampling, image segmentation, feature extraction, etc., and it combines computer vision, multi-dimensional signal processing and numerical calculation closely. Image registration technology is closely related to the research directions of image fusion and image segmentation, and is the research foundation in the fields of image understanding and image restoration.
According to the different image information used in image registration, image registration methods can be divided into three main categories: gray-level information-based methods, transform domain-based methods, and feature-based methods. It can be subdivided into several categories according to the selected characteristic attributes.
·Image registration method based on gray information
This type of method generally does not require complex pre-processing of the image, but uses the grayscale statistical information of the image itself to measure the similarity of the image. Its main feature is that it is simple to implement, but has a narrow application range, and cannot be directly used to correct the non-linear deformation of the image. It often requires a huge amount of calculation in the search process of the optimal transformation.
·Image registration method based on transform domain
The most important transform domain method is the Fourier transform method. This method takes advantage of the good properties of the Fourier transform, that is, the function translation, rotation and scaling have their symmetry in the frequency domain. For the image translation, calculating the Fourier transform of the power spectrum of the two images can obtain an impulse function, which is not zero only at the amount of translation. For rotation, it can be expressed in polar coordinates, so that the rotation of the image is transformed into the translation of the image, and then the rotation angle between the images is calculated in the same way. If there are not only translation transformations but also rotation transformations between the images, we calculate in two steps: first calculate the rotation transformation and then calculate the translation transformation. This method is very suitable for image registration with small translation and rotation and scaling. At the same time, it has hardware support and fast algorithms, so the calculation speed is fast, and it can overcome correlated noise and frequency-dependent noise, and is suitable for images collected by multiple sensors and light source changes.
·Feature-based image registration method
Feature-based image registration methods are a large category of image registration methods. The main commonality of these methods is that the image to be registered must be preprocessed first, which is the process of feature extraction, and then the extracted features are used. Complete the matching between the features of the two images, and establish the registration mapping transformation between the images through the matching relationship of the features. Because there are many available features in the images, a variety of feature-based methods have been produced. Commonly used image features include: feature points (including corner points, high curvature points, etc.), straight line segments, edges, contours, closed regions, feature structures, and statistical features such as moment invariants, center of gravity, and so on.
In this topic, we will focus on the feature-based image registration method.
1.3 Application of image registration in medicine
Medical image registration and medical image fusion are closely related, especially for multi-modal images, registration and fusion are inseparable. The images to be fused often come from different imaging devices, and their imaging orientation, angle, resolution and other factors are different. Therefore, the positions and sizes of the corresponding tissues in these images are different. If the fusion image is not spatially adjusted in advance Alignment on the upper, then the fused image is meaningless. Therefore, image registration is a prerequisite for image fusion, and registration transformation must be performed first to achieve accurate fusion.
Medical image registration is a basic task of medical image processing. It can register multiple images from different modalities or at different times, and then provide guarantee for image post-processing. For example, in medical image fusion, the corresponding tissue structures need to be fused together, and the images to be fused often come from different imaging devices, and their imaging orientation, angle, resolution and other factors are different, so these images The location and size of the corresponding organizations in the system are different, and the registration transformation must be performed first to achieve accurate integration. The definition of medical image registration: Two photos taken by the same person from different angles and different positions. Due to different shooting conditions, each photo only reflects certain aspects of the characteristics. To analyze the two photos together, you must move and rotate the portrait in one to align it with the other. This alignment process is the registration process. The image that remains still is called the reference image, and the image that is transformed is called the floating image. By fusing the registered images, a fused image reflecting the whole picture of the person can be obtained. Medical image registration is to seek the geometric transformation relationship between two images. Through this geometric transformation, the corresponding points on one medical image (floating image F) and the other medical image (reference image R) can be spatially reached. Consistent. This consistency means that the same anatomical point on the human body has the same spatial position on the two matching images. The result of registration should match all anatomical points on the two images, or at least all points of diagnostic significance and points of interest for surgery.
1.4 issues with the software Matlab Introduction
MATLAB , the Matrix Laboratory for short, is the United States Mathworks company in 1984 launched the annual numerical computer simulation software , through continuous development and improvement , now has become a cover multiple disciplines , is a powerful software has the ability and numerical simulation analysis capabilities. The application of Matlab is relatively simple . It uses mathematical expressions that everyone is very familiar with to express problems and solving methods. It integrates calculation, graphics and programming into one environment, which is very convenient to use. At the same time , Matlab has strong openness and adaptability . While keeping the kernel unchanged , Matlab has launched toolboxes suitable for different disciplines , such as image processing toolbox , wavelet analysis toolbox, signal processing toolbox, neural network Toolboxes, etc. , greatly facilitate the research work of different disciplines. Matlab powerful drawing capabilities , a simple form of the command , making it more and more popular scientific and technical personnel at home and abroad , to be more widely applied[8]。
The reason why MATLAB is popularized so quickly and shows such vigorous vitality is because it has characteristics different from other languages. Just as high-level languages such as Fortran and C free people from the need to directly manipulate computer hardware resources, MATLAB , known as the fourth- generation computer language , uses its rich function resources to free programmers from cumbersome program codes. come out. The main features of MATLAB :
·Powerful
MATLAB has a powerful toolbox, which contains two parts: the core part and various optional toolboxes. In the core part, there are hundreds of core internal functions. The toolbox can be divided into two categories: functional toolbox and discipline toolbox. The functional toolbox is mainly used to expand its symbolic calculation functions, graphic modeling and simulation functions, word processing functions, and real-time interaction with hardware. Functional toolboxes can be used in a variety of disciplines, and subject-based toolboxes are relatively professional, such as control toolbox, image processing toolbox, signal processing toolbox, etc. These toolboxes are written by experts with high academic levels in the field, so users can directly conduct high-level, precise, and cutting-edge research without writing basic programs within their own disciplines. The reason why MATLAB has become the world's top scientific computing and mathematics application software is because it has more and more powerful functions with the upgrade and continuous improvement of the version [9] , mainly including: numerical calculation function; symbolic calculation function; data analysis Function; dynamic simulation function; unified graphics and text processing function.
·Friendly interface, high programming efficiency
The outstanding feature of MATLAB is simplicity. It replaces the verbose codes of C and Fortran with more intuitive codes that conform to human thinking habits , and brings users the most intuitive and concise program development environment. MATLAB language is concise and compact, easy to use and flexible, library functions are extremely rich, program writing forms are free, use its rich library functions to avoid complicated subroutine programming tasks, and compress all unnecessary programming work. Since library functions are written by experts in the field, users do not need to worry about the reliability of the functions. It can be said that scientific development with MATLAB is done on the shoulders of experts.
·Strong openness
MATLAB is very extensible and can be used as a higher-level language. The functions in the various toolboxes can call each other or be changed by the user. MATLAB supports users to carry out secondary development of their functions, and users' applications can be added to the corresponding toolbox as new functions.
Image registration is the process of matching and superimposing two or more images acquired at different times, different sensors (imaging equipment) or under different conditions (weather, illuminance, camera position and angle, etc.). It has been widely used In the fields of remote sensing data analysis, computer vision, image processing, etc.
The process of the registration technology is as follows: first extract feature points from two images; find matching feature point pairs through similarity measurement; then obtain image space coordinate transformation parameters through the matched feature point pairs: finally coordinate transformation parameters Perform image registration. Feature extraction is the key to the registration technology, and accurate feature extraction provides a guarantee for the success of feature matching.
Therefore, the key problem to be solved in this subject is to design an algorithm to match two or more images acquired at different times, different sensors (imaging equipment) or under different conditions (weather, illuminance, camera position and angle, etc.). Overlay.
2.2 Research methods of feature-based image registration
2.2.1 Image preprocessing
For the reference image and the image to be registered, since the images are not captured at the same time or the same sensor, in order to eliminate the gray difference between the reference image and the image to be registered as much as possible, we generally choose histogram matching as the preprocessing step. The histogram matching processing technology uses the group mapping rule GML to realize the mapping between gray levels.
2.2.2 Feature point extraction
Feature-based methods often extract obvious regional blocks, line structures and key points in the image as features. These features are required to be significant enough to be easily detected under various distortion conditions. Compared with the cross-correlation algorithm that directly uses the pixel gray information, the feature extraction contains high-level signal information, so this type of algorithm has stronger anti-interference ability against light and noise.
In this topic, we will focus on the image registration of point features. On the same image, the Harris corner points, SUSAN corner points and SIFT feature points are extracted from the reference image and the image to be registered after the histogram is matched.
2.2.3 Choose a matching strategy
Different feature points can try different matching strategies. This topic will adopt the following methods for experimentation:
·The Harris corner-based matching strategy adopts cross-correlation method for rough matching, and then uses virtual triangles for precise matching;
·The matching strategy based on SUSAN corners adopts the cross-correlation method for rough matching, and then uses the RANSAC strategy for precise matching;
· The matching strategy based on SIFT feature points is roughly matched by the Euclidean distance of the feature vector of each point, and the RANSAC strategy is used for precise matching.