Acceleration Algorithm Research For Image High Dimensional Feature Matching

In recent years,the image retrieval technology is developing rapidly,and the image retrieval application based on GIST features and SIFT features is increasing year by year.The general feature of the image is extracted from the specific image,it is a kind of representation of the image information,and the high dimensional feature has the characteristic of floating point type,which has the characteristics of the high dimension.Some of these binary features are extracted from the image,some of which are derived from the simplified reduction of floating point features,which are more advantageous than floating point features.Nowadays,the characteristics of high latitude restrict the speed of image search,in this case the most important question is how to simplify the feature of image retrieval speed or improve the accuracy of image retrieval.We study the floating point features,the traditional method is to reduce the dimension,but it also reduces the accuracy of image retrieval.In this paper,we try to find a way to improve the speed of image retrieval without reducing or decreasing the precision of image retrieval.The content of this dissertation is as following:1.In this dissertation,we propose a new image retrieval algorithm.We studied the existing methods,moved the calculation of the distance between the image and image database t off line,ranking the images off line.When the user searches,they only need to calculate a few function values to find the closest image.According to the idea of the algorithm,we have respectively carried out the image retrieval based on GIST and SIFT features.In the design of an pre-ranking image retrieval algorithm based on GIST feature,this paper designs two fitting functions according to the size of the cifar10 database.In order to increase the accuracy of image retrieval,this paper designs a small range search function Find().In the experiment,the pre-ranking image retrieval algorithm based on GIST feature is compared with the traditional algorithm of PCA dimension reduction+ITQ dimension reduction and nearest neighbor retrieval.The experimental results show that the proposed algorithm can improve the retrieval speed and accuracy of image retrieval with GIST feature when the cifar10 data set is used.In the design of an pre-ranking image retrieval algorithm based on SIFT feature,because of the large number of SIFT features extracted from the Dup Image data set,four fitting functions are designed in this paper.In order to increase the accuracy of image retrieval,the similarity degree of the image is normalized and relative similarity is obtained.We compare an pre-ranking image retrieval algorithm and BOW algorithm based on SIFT feature.The experimental results show that the accuracy of image retrieval is improved by 6.76% when thealgorithm is applied to image retrieval with SIFT feature.The Dup Image algorithm is an excellent and feasible algorithm for image retrieval.2.In this paper,we propose a k-d tree matching search algorithm,which is an alternative to the nearest neighbor search algorithm without reducing the search accuracy.In this paper,we propose a scheme to replace the nearest neighbor search algorithm without reducing the search precision,We put the binary features in a certain order stored in the tree structure.When matching we search the tree.It greatly speeds up the speed of matching search.When it is realized,it includes the r-nearest neighbor search algorithm at low dimensional and the r-nearest neighbor search algorithm at high dimensional.Two algorithms build two trees in different ways.The low dimensional algorithm and the nearest neighbor search are compared in the experiment.The data set use cifar10,and the experimental data show that the average running time is reduced by1.14 seconds.The r-nearest neighbor search algorithm at high dimensional is compared with the linear search algorithm.The experimental results show that the average running time is reduced by 1.71 seconds by using the cifar10 data set.They can be used to accelerate the search speed in different situations.