A Study On The Technology Of Image Retrieval Based On Android

Image retrieval is a very important job in computer vision research. With therapid development of the popularity of smart phones and mobile Internet, there is asharp rise in the demand for mobility and multimedia information. Through theeffective integration of smart phones and image retrieval technology, people canobtain visual object from real-world and interest relevant information from the otherend of the mobile Internet quickly and easily.This paper first introduces the research background, research actuality, challengeof photographed image retrieval and the processes of photographed image retrieval onphone. Then this paper proposes two photographed image retrieval algorithm andimplements a photographed image retrieval system by studying various local imagefeature extraction methods, feature descriptors, feature matching and image retrievalalgorithm deeply. Main contributions are as follows:1. The classic feature point extractions and descriptor calculations are performedon a gray scale, however for image local regions with a similar structure but differentcolors, they can produce false matching result. For the images which have moresimilar structure contents, there will be more mismatching. To address this problem,this paper proposes a photographed image retrieval algorithm based on ORB feature.By setting threshold to increase the initial ORB matching rate. Thus, the increasedcorrect matching pairs meet homography mapping, and the increased error matchingpairs are randomly. Therefore the correct matching pairs which meet homographymapping is the most and improve RANSAC homography calculation accuracy.Finally use the image color information to exclude false positive matching pairs. ORBfeature is very suitable for mobile phones because it has a rapid calculation andoccupies less memory. Experimental results show that the proposed photographedimage retrieval algorithm is robust in geometric attacks, signal processing attacks andattacks from photographing.2. By comparing the final matching rates in different classic feature pointextraction methods on photographed images, we can find that Harris method has abetter match rate and stability. However, Harris method has no scale information,which is not suitable for photographed images, because photographed image is ofteninconsistent with the resolution of the original image. To address this problem, this paper studies the scale space theory and proposes a Harris algorithm with theinformation of spatial scales. The proposed algorithm can be used in photographedimages. Experimental results show that the multi-scale Harris algorithm has a highmatching rate in mapping different resolution photographed images.3. This paper implements a photographed image retrieval system based on theproposed multi-scale Harris algorithm and the improved matching algorithm. Thissystem includes two parts, one is image acquisition, feature and descriptors extractionon the end of phone, another is image feature matching and retrieval on the server.The system is experimented in this paper. Experimental results show that theimplemented photographed image retrieval system meet all the demands for use, has ahigh identify rate, fast speed and is robust.