Study Of An Image Retrieval Method Based On Content Self-Organized And Interactive

With the rapid development of the multimedia network technology, theapplication of the image becomes more and more extensive. And the use of theinformation of images has come into the every walk of life, the image has been oneof the most forms of the popular digital information. It is very important to manageand search the image resources. The study of the image database will providewith the development of the multimedia digital library, medical image management,satellite and remote sensing image and the computer aided design and manufacture,geography information system, criminal identify system and logo copyrightmanagement.One of the kernel technology of the study of the image database is how tolookup the image you need and make it very convenient, available ,fast andveracious, and it becomes the choke point of the management of the tremendousinformation. The Content-based Image Retrieval (CBIR) has been presented andbecomes one of the most active researches in multimedia retrieval field.Content-based image retrieval (CBIR) system is different from traditionalretrieval system based on text mode. Index of image is described by some visualfeatures but not by text, such as, color, texture, shape, and locality feature etc.Users always search similar image according to an example image. Result ofretrieval can be feedback with ranked similarity degree, and similarity degree iscomputed from distance function in feature vector space.In this paper, we extensively studied the national and international materials onCBIR systems, discuss the research status and trend in content based imageretrieval problems, and from the investigation we find: because of the variety of theretrieval objects and the range, the study of the CBIR has a far-ranging content.Though there have been a good many of CBIR algorithms and have a few retrievalprototype systems, the retrieval precision always needs to be investigated.Furthermore, letting the system and the user learn each other in the course of theretrieval alternation, catch on the content and let the retrieval performances beclosed to the human visual characteristics well are also the questions to be solvedin the study. In this paper, for the research and the development of the field, wedesign and implement a content-based self-organized and interactive retrievalsystem.With the development of the epoch and the advancement of the technology,image resources become more and more ample and the image database becomesmore tremendous, then for users, it is inconvenient to use the database. Then howcan we find the image we need in the tremendous database expediently? Then wefind it is a good method to self-organize the image database based on the content,and the image database change from out-of-order to regularity. That is similar toour Jilin University divided into different institute, different department anddifferent class, according to the rule, students are divided into idiographic institute,idiographic department and idiographic class. In the case of that there is nopartition, how we can do when a paterfamilias want to see her sun peremptorily? Itis too difficult to find him in the way of one-by-one in our large number of students.At the same time, if we search for him according to the dividing rule, we'll find thestudent easily. We can introduce the method to the management of the imagedatabase, and the only difference is the dividing standard. As we all know that inorder to analyze the information of an image, the CBIR system always analyzes thecolor, texture, shape, and other low-level image features. In our dissertation, we usethe similarity of the spatial relationship feature and then classify the database intodifferent sorts. At the mention of classifying, we find Fuzzy C-means (FCM)algorithm is used as a clustering algorithm broadly and effectively now. And one ofthe most keys of the FCM is the confirmation of the fuzzy weight. The popular wayto calculate the weight is using the distance between the feature vectors, but whenthe distance between a vector and one centroid is equal to the other, no matter howthe two clusters distributing, the weights are equal, then we can't get exactclustering. For the sake of solving the problem and well reflecting the clusterdistribution, we adopt a new way to calculate the weight.. And then the subjectionof the image cluster can be reflected exactly. After the course of theself-organization based on the similarity of the content, the image databasebecomes regularity. When we want to find the image we need, we can find thecentroid of the smallest distance between the sample and cluster centroids firstly,and then find the image in a cluster, that is ok. It is much more efficient thanfinding order by order in database, too.For the image database, after self-organized based on the similarity of thecontent, the convenience of the use of the database has been enhanced awfully, butthe precision descended a certain extent. We couldn't attend to one thing and loseanother, so we must consider both the efficiency and precision.At present, the CBIR system always analyzes the color, texture, shape, andother low-level image features, though these low-level visual features apart fromthe high-level semantic features. Low-level features can't distinguish the objectof the image. Therefore, no matter what features you adopt and no matter whatdistance function you employ, two images resembled whether or not is rely on theuser ultimately. So we deem that the user not the computer is the center of theCBIR system. In addition, owing to the user's different emphases, the standard ofthe similarity is diverse largely. How to let the system be capable of adapting thesedemands automatically should be researched as soon as possible, then we can getbetter query results and the query precision enhanced in the same way. It isimportant to add the user's subjective like and sensibility to the CBIR system, andcarry through interactive query (IQ), the course of the IQ is also named relevancefeedback (RF) technique.For the sake of improving the query precision and implementing theapplication of the RF in the CBIR, we take two measures to improve theperformance of CBIR in the dissertation. For one thing is the account of thesimilarity, most systems are always using unitary method now, for example EuclidDistance, but in nature, not all of the images are adapt to the method you adopt, sowe bring forward a dynamic selective distance function in the paper. The userselect the relevant images that he think based on the first query result, then therelevant images return to the system, and the system makes use of rank operationand select the best distance function as the similarity standard, finally return theresult by using the best distance function. For another, the system adjusts the queryvector dynamically based on the user's feedback. The method described as follows:calculate the distance between the sample and the relevant images by the spatialrelationship and color feature respectively, then select the maximum distancesindependently, one is for the spatial relationship feature, the other is for the colorfeature, calculate the sum of the two distances, and the two features weights areconformed by the inverse proportion, at last the system goes along a new queryaccording to the weights. After the course of the interactive query, the precision isheightened a certain extent, and that fetches up the deficiency of the clustering forthe image database synchronously.In this dissertation, we mend the system of our laboratory and the system fordeveloping this is Windows xp, and the development environment is Visual C++6.0 and Microsoft SQL Server. This dissertation holds certain referential value andpractical significance in promoting the development of retrieval technique of imagedatabase.