Image Retrieval Algorithm Based On Feature Matching

In today’s information technology rapid development, digital cameras, cell phones, thedramatic increase in age of the monitoring equipment, digital image, with its large amount ofinformation and the great advantage of the easy-to-understand, has been applied to the societyin all walks of life, such as industrial manufacturing, media, traffic monitoring, health andfamily life. However, while digital images bring convenience to our daily life, how effectivemanagement of a large number of digital image database, and be able to quickly and accuratepositioning of the images from the mass of the target image, become an urgent need to solve,so looking for an efficient image retrieval techniques became the target.To solve this problem, this paper first introduces the existing image retrieval technology,and elaborates insufficient of traditional text-based image retrieval, then, we focus oncontent-based image retrieval technology. This paper mainly analyzes the globalcharacterization method and the description method of local feature points. By analyzing theiradvantages and disadvantages, we selected the local invariant features SIFT (scale invariantfeature transform).we studied the principles and extraction process of the classical localinvariant features–SIFT in Chapter III, and carried out some experiments using SIFTfeatures. The experimental results showed that the SIFT features had good validity in terms ofobjects recognition, but their extracting time was so long that they could not be used in apractical application. As a result, we did a depth research on the improved algorithm of SIFT,we introduced the K-D tree algorithm to index all the feature points extracted from thedatabase images in the tree structure. After that, we used the K-D tree search algorithm basedon BBF to find nearest neighbor and second nearest neighbor of the query feature point in thebuilt K-D tree, which improved the image matching efficiency. Thirdly, we combinedRANSAC algorithm to do the consistency test on the successfully matched feature points. Thealgorithm randomly created a number of mathematical models and calculated thecorresponding numbers of interior points. Then it selected the optimal mathematical modelwith the largest interior point’s number and used it to remove the error match between images,which improved the accuracy of feature matching. Finally, we introduced PCA-SIFT–theimproved algorithm of SIFT, it made SIFT128dimensional feature vector reduced to36dimensional to improve the speed of matching.Based on the proposed image retrieval method, we constructed a fast and robust imageretrieval system. The system consisted of two modules of the offline processing and onlineprocessing, which were responsible for processing the image database and query image recognition request respectively. Subsequently, we used parts of images to do the validationexperiments on the established system. The results showed that the constructed imageretrieval system fitted all the technical requirements with high recognition accuracy andrunning speed, which was a fast and robust system.