Research Of Image Retrieval Technique Based On SIFT

Along with the large image databases, digital libraries, network expansion, thecontent-based image retrieval technology is more and more frequently appeared inpeople’s life, large number of researchers pay attention of it. Content-based imageretrieval technology different from the text information retrieval, it don’t require thetedious manual identification such as "index","catalog" and "abstract", but use of thevisual information of image content, from low-level to high-level automatic layers ofprocessing, search of the image features to complete the matching.This article in view of the new requirements for the innovative and exploratorystudy on the content-based image retrieval technology.The main research work isdivided into the following several aspects of content:(1)According to the theory of content-based image retrieval technology， in termsof local characteristics, due to the SIFT feature has superior retrieval accuracy, thispaper adopts the SIFT algorithm feature extraction and image descriptor toidentification the different scenarios, the experiment proved its remarkable scaleinvariance and rotation invariance;(2)Study deeply in the similarity measurement method. Put forward combined withdistance measurement and statistical measurement of the similarity calculation. At thistime in order to improve the matching efficiency in this paper, adopted the BBF-KDtree algorithm to match the feature points of image processing. At the same time weuse RANSAC algorithm to eliminate the error information, through a large number ofexperiments to verify the effectiveness of this algorithm is superior to other effects.(3)To improve the spatial visualization dictionary model, combined with the Bag-of-Feature model and the SVM algorithm, using the image pyramid block in Bag-of-Feature model, using the SVM algorithm to matches every word bags. Through thesimulation experiments we can found that the new algorithm has good performancesin precision and recall rate.To sum up, the system of content-based image retrieval technology is aimed at therequirement of accurate retrieval using SIFT algorithm for image feature extractionand image descriptor to identification, combined with distance measurement andstatistical measurement in similarity calculation, adopted the BBF-KD tree algorithmto match the feature points during the image processing, and put forward theimprovement of spatial visualization dictionary model. And verify the algorithm’s theeffectiveness and accuracy through the simulation experiments.