Research On Extraction And Matching Of Image Local Features

Compared with global features,local features are not easily disturbed by illumination,occlusion and other factors.At the same time,local features are more accurate and robust for image representation and description,and can effectively describe image content for image retrieval and target recognition.Local features mainly include direct gray descriptors and feature point algorithms based on local regions.Local feature algorithms based on feature points are proposed,which include corner points and gray extreme points.These algorithms are basically composed of three parts: feature point extraction,feature point description and feature point matching.The different algorithms of these three parts have great extensibility.The combination of different algorithms can improve the local feature algorithm.For the main corner algorithm,because of the fixed operator scope,it does not have scale invariance,so the adaptive operator scope can effectively improve the scale adaptability of corner operator.The main contents and innovations of this paper are as follows:(1)Firstly,different corner algorithms,including Harris corner,SUSAN algorithm,FAST algorithm and local feature matching algorithm,including SIFT/SURF and ORB,are deeply studied.These algorithms are classified and summarized according to the three parts of feature point extraction,description and matching.This paper summarizes the three steps of the current local feature algorithm process,studies the ideas of different local feature algorithms,sums up the concept of modularization of local feature algorithm,summarizes the modularization method,combines the different method modules crosswise,carries out comparative experiments under scale,rotation and perspective transformation,and carries out theoretical analysis on the performance of different method modules.It evaluates the advantages and disadvantages and extensibility of different modules,and summarizes the improvement space of different algorithms.(2)Aiming at the problems of too long multi-scale analysis flow and low efficiency of image feature algorithm,a scale information entropy model is proposed from the perspective of detecting relative scale of image.Based on the relationship between information entropy and image gradient in scale space,a mathematical model is established,and a large number of experiments are carried out to fit the model.Finally,the functional model of scale information entropy model is obtained.According to the relative scale parameters between the images to be matched,image restoration,i.e.deblurring operation,is performed on the high-scale images to obtain low-scale images.The effectiveness of the scale information entropy model in this paper is verified by the inverse process.At the same time,a multi-scale improvement method for FAST corner points is proposed.Based on the scale information entropy model,a multiscale FAST corner point with adaptive scope is proposed.The comparison experiment proves that the improved multi-scale FAST corner points in this paper have good scale adaptability.Finally,this paper describes the application and implementation of local feature algorithm in image mosaic,analyses the basic mosaic process,and carries out panoramic mosaic experiments with experimental images collected by intelligent devices.