A CBIR Algorithm Based On The Adaptive Adjustment Of Feature Weights

With the development of multimedia and internet technology, imagehas been an important information source besides text. Imageretrieval technique comes after the need for management of the imageresources. We can use traditional text indexing method to retrievethe images; however, the text indexing technique cannot make full useof the rich information of the images. Image retrieval can usehigh-level semantic based techniques, but it has not reachedapplication level. Content based image retrieval technique has beenthe research focus and key technology of current informationretrieval research. CBIR refers to the image indexing techniques bythe content of the image. Its main idea is based on the analysis ofthe color, texture, shape and object spatial relationship embeddedin the image. Index of the image is built on the feature vector andthe current retrieval methods are based on the similarity matchingof multi-dimensional vectors.Because of the objectivity of people's vision, especially for thebig gap between the high-level semantic meaning and the low-levelimage feature, we cannot get the satisfying map relationship throughthe existing computer vision and artificial intelligence techniques.So we can only use feature extraction, however, it can result goodperformance. In order to solve this problem, relevance feedback (RF)techniques have attained wide application.In the content based image and video retrieval, feedback modifiesthe information to adapt to the client's preference and tune theretrieval performance. Especially in the retrieval through semanticmeaning, feedback realizes the interoperation between people and themachine; hence it involves people's knowledge into the retrieval.Most common and more mature feedback methods are relevancefeedback. People modifies their indexing requirement according to theprevious retrieval results, in this way the system can provide betterresults to satisfy people's requirement. Relevance feedback is a kind of interoperate indexing method. Inthis process, the clients are asked to give the judgment over thecurrent retrieval results and the system re-studies the clients'feedback. So the system can understand the clients'requirement better,and gives better performance. In CBIR, relevance feedback works asfollows: First, the clients specify the sample image and submit itfor indexing. The system performs the similarity matching between thefeatures of the sample image and the features of the image in thelibrary. A list of the similarity degree is computed, in which thecandidate retrieval images are sorted according to the similaritydegree with the higher degree in the higher position. Then the clientschoose a group positive candidate images that satisfies the retrievalrequest or a group of negative candidate images that do not satisfythe retrieval request from the library and submit them to the system.The system modifies the weight or the optimum index vector accordingto the clients'feedback and provides the new retrieval results. Sothe most important issue in the relevance feedback is how to make fulluse of the feedback information provided by the clients and modifiesor optimums the index vector to better the retrieval performance. Feature weight modification is a kind of relevance feedbackalgorithms. The basic idea of feature weight modification is toenhance the feature weights that are helpful for the image retrievaland at the same time reduce the negative ones that are not helpfulfor the image retrieval. MARS system realizes the progress of theoff-standard weight modification method. Weight modification methodstry to modify each feature and the weight of each dimension of thefeature to optimum the retrieval performance. A feature standardvariance based method was proposed. Its basic idea is setting lowerweight to the features with bigger feature standard variance, on theopposite side, setting the higher weights. In the same time theyprovide a multi-layer image model and based on this model they usethe global optimum method to solve the threshold modification issue. This thesis is carried out through mass of national andinternational relevance feedback materials and the current trend andresearch status of relevance feedback techniques in image retrievalissues has been given in it. We studied the main techniques about CBIR,conclude the current relevance feedback techniques and present theproposed relevance feedback algorithm. The algorithm aims at theimportance of the function of relevance feedback techniques on theCBIR system and presents the idea of the combination of the low-levelfeature and the high-level semantic meaning of the image, therelevance feedback and the machine learning. Make full use of theprocessing ability of the computer system and the clients'feedbackinformation and better the retrieval performance of the CBIR system. The algorithm of this thesis makes use of the database techniquesand get the semantic information to store in the memory through theinteroperate feedback. This stored information can better theretrieval performance in the following retrieval process. Theattained and the stored feedback information is a results set of theouter weight and the clients'preference in the two feature space.They can act as the initial condition of the successive indexing workand measure of the indexing results. The derivative of the resultsset is a machine learning process and it can be realized through thefeedback between the clients and the system. To modify the featureweights, the standard is set according to the results set. The systemdetermines the weight modification performance and further modifyingthe weights or the retrieval results. The retrieval results throughthe interoperation of the clients and the system can perfect theoriginal retrieval performance. In this thesis we test theprogressive method, the reduce-by-half method and the golden...