Research On Image Registration And Image Stabilization Based On Feature Point Matching

Image registration technology is one of the key processes in the application of image stitching and target recognition and tracking.Due to its wide application,it has been studied all the time.Image registration methods includes two categories: region-based image matching methods and feature point-based matching methods.The feature point-based method analyzes important information of the image,so it can greatly reduce the amount of computer work in the matching process.In addition,the feature extraction can resist noise and can better adapt to changes such as grayscale,viewing angle,and occlusion.Therefore,in this paper,the registration method based on feature points is selected for research.With many years of research by home and abroad scholars,great progress has been made in image registration and its related applications.However,for scene images with too many dynamic changes,illumination transformation and similar features,there is a certain gap between the accuracy and real-time accuracy.In addition,the matching method based on feature points and motion filtering are applied to the processing of video sequences to improve the stability of the video.There are three main research methods:(1)In terms of feature extraction,an optimal threshold prediction method for image feature extraction under mixed features is proposed.Aiming at the problem of low image matching rate under single feature conditions and the uneven extraction of feature points due to fixed contrast threshold of SIFT(Scale-invariant feature transform)algorithm.The improved algorithm in this paper first uses SIFT algorithm to extract image feature points,then uses the characteristic second-order moment to construct the threshold prediction model,and then uses highly descriptive texture feature vectors to constrain the SIFT matching process to achieve image matching.The experimental results show that the improved method adaptively adjusts the contrast threshold according to the gray distribution of the image,which can stabilize the number of extracted feature points and can effectively reduce the phenomenon of similar mismatching.(2)For the matching of sequence images,a new adaptive threshold matching method based on sequence images is proposed.When SIFT deals with sequence image matching in complex environments such as lighting and blur,due to the lack of consideration of specific features of the scene by setting a fixed threshold in advance,the matching difficulty and low matching rate will occur.To solve these problems,the improved method of this paper first analyzes the statistical characteristics of sequence images,then calculates the comprehensive indicators of each statistical characteristic according to the principal component analysis method,and establishes an adaptive threshold prediction model for feature point detection based on the main influence factors and statistical characteristics.Experimental results show that an adaptive threshold prediction model overcomes a variety of extreme conditions and improves the matching effect of sequence images in complex environments,especially when processing illumination sequence images.(3)After researching and analyzing the principles and methods of electronic image stabilization processing,a registration method based on feature matching and motion filtering are applied to electronic image stabilization in this paper.Firstly,the model estimation parameters between images are calculated by the matching method based on feature,and then the random dither vectors are eliminated by kalman filter to correct and compensate the parameters.The results show that the random jitter vector in the video can be effectively removed after the motion filtering process,and the stable scene video can be effectively output.