Research On Image Registration Method Based On Local Point Feature Extraction

In recent years,with the rapid development of computer vision technology,image registration as the most basic problem has gradually become a popular research technology in the field of image processing.The technology combines image preprocessing,feature extraction,similarity metrics,grayscale interpolation and other related knowledge.The ultimate goal is to achieve optimal registration of images acquired from different sensors,different time and different perspectives.At present,image registration technology has been applied to many fields,such as image data fusion,image difference detection,three-dimensional image modeling,and pattern recognition.At present,feature-based image registration is one of the most valuable methods in many image registration methods.This method first extracts the common features from the image as the registration primitives to establish the correspondence;Then solves the transformation model parameters and transforms the image;Finally,performs gray interpolation on the non-integer pixels of the registered image,complete the final registration.The accuracy,stability and timeliness of the algorithm will directly affect the quality of the final result.In recent years,although many scholars have carried out research on it and made great breakthroughs,according to practical application considerations,there are still some unsolved practical problems: How to find a balance between algorithm accuracy and computational efficiency,so that it can adapt to the real-time nature of the image;Whether it can extract the image features with high utilization rate,and facilitate the subsequent feature matching.Through the in-depth analysis and summary of the advantages and disadvantages of feature extraction and similarity measurement methods in image registration methods,based on the research of predecessors,this paper proposes some improvement ideas for some problems.The specific work is as follows:1)For feature extraction,an adaptive Harris corner detection algorithm based on template edge is proposed.Aiming at the problem that the traditional corner detection algorithm needs to perform gray operation on the whole image,before extracting the corner point,the paper extracts the potential corner area by the idea of local area,which can effectively improve the efficiency of the subsequent corner detection algorithm.Aiming at the problem that the corner points extracted by Harris corner detection algorithm have pseudo corner points and threshold artificial settings,the threshold setting is improved to change according to the gray level change of the potential corner area,and a circular template edge model is proposed.This model is used to perform secondary detection on the extracted corner points,and the pseudo corner points and cluster corner points are eliminated.The experimental results show that the improved algorithm can effectively filter out the pseudo corner points,detect the corner points with obvious attributes,and improve the correct rate of corner point extraction.At the same time,the time used in the experiment was compared,and the calculation efficiency was greatly improved compared with the traditional Harris algorithm.2)For the similarity measure,an image registration method based on SIFT(Scale Invariant Feature Transform)and improved RANSAC(Random Sample Consensus)is proposed.Firstly,the corner point extraction algorithm is used to extract the corners of the image.The Descriptor Extractor in the SIFT algorithm is used to determine the main direction and feature description of the extracted corner points to obtain multi-dimensional feature vectors.Then,the Euclidean distance of the feature point feature vector is used as the similarity determination metric to complete the rough matching of the feature point pairs.There is a pair of error points in the feature point pairs obtained after the rough matching.Most scholars will use the RANSAC algorithm to eliminate them.In this paper,the problem that the RANSAC algorithm needs to be further refined and the parameters need to be refined when selecting the subset is selected.The improved is used Randomized RANSAC(R-RANSAC)with Sequential Probability Ratio Test(SPRT)algorithm to eliminate the mismatched pair and is purified according to the matching point pair calculates the geometric transformation parameters of the two images,inversely transforms the image and interpolates the grayscale to complete the image registration.The simulation experiment is carried out with randomly selected images with moderate complexity.The results show that the improved algorithm can achieve faster matching speed under the premise of ensuring the accuracy of registration.The time comparison between RANSAC algorithm and R-RANSAC algorithm is carried out.The calculation efficiency of the latter is significantly higher than the former.