Study On Key Issues For Multimedia Retrieval Based On Semantics

With the development of network, a lot of digital multimedia information increases surprisingly and people's requirement for many kinds of digital multimedia increases rapidly, then the problem of how to searching multimedia with no wasting time and locating exactly what the users want appears among the large numbers of multimedia. The important research aspect which is how to efficiently organize, manage, search multimedia data appears.The paper discusses the problem of searching multimedia data based on semantic information, mainly about: the paper analyzes multimedia information by multi-dimension, by which gives aggregated experiment based on multi-dimension features of image's semantics; The paper gives mutli-dimension storing structure for semi-structure multimedia data, by which proposes weight set automatically algorithm based on multi-dimension features and improve the retrieval efficiency; The paper gives high-level semantic mathimatical model for efficient organizing high-level semantic information, by which proposes semantic branch structure for video key frame. The author studies the multimedia retrieval from different semantic angles. According to the current status of searching multimedia, the paper gives how to describe multimedia by multi-dimension, with which the researcher can analyze and operate multimedia information (image, video, audio) from different dimensions. Through describing multimedia by multi-dimesion, we may combine many kinds of multimedia efficiently and locate multimedia information exactly by one dimension or several dimensions and analize multimedia information from different angles: different categories, different features and different semantic level. Accordingly we can carry out the mechanism for searching multimedia information by multi-dimension. The paper combinds two kind of classifier AdaBoost with SVM and makes SVM be the weak classifier for Adaboost, the weak classifier is linear kernel SVM in the paper. The features for this kind of SVM are that the parameter is simple, operating speed is quick and easy using. It is more important that the iterative process increases the classifier precision in itself, so the weak classifier does not need high precision with combinding AdaBoost with SVM. When classifier is designed, the weak classifier, namely SVM classifier, is not requested that the internal parameter C is fixed. The C parameter may be adjusted conveniently to increase otherness of the weak classifiers for realizing iterative process optimization every iterated time. The experiment shows that the result is satisfied with SVM-AdaBoost aggregated algorithm based on multi-dimension featuresIt is that the relational model is fit for structural data, but it is not fit for the semi-structure data—multimedia. The paper proposes one storing structure by multi-dimesion which is fit for multimedia information. We can store multimedia information with different angles: different categories, different features and different semantic level, then establish index, which is analyzed and stored in the memory before retrieval. The popurse is improving the retrieval efficiency. The methods of operating and analyzing for multi-dimesion is fit for the storing structure, then we uses the same methods used in multi-dimension multimedia for the storing structure, which we can analyze and operate from different dimensions. Thus, we can combind multi-dimension multimedia with multi-dimension storing structure, then we can obtain better effect with small price. The paper proposes one step-sized statistics algorithm based on multi-dimension feature for the problem of setting multi-dimension features'weight automaticly based on the multi-dimension storing structure. When users submit different images for request, we can set every dimension feature automatically as we want by the algorithm. We make the weights of different features, such as color and texture summative for different feature vectors to achieve the purpose of multi-dimension retrieval. So we can avoid a large number of subjective factors of man-made intervention and time waste.Because the step-size statistic algorithm based on multi-dimension feature limits the scope of step, we involve genetic algorithm and propose the weight set automatically algorithm based on multi-dimension feature. We set every dimension feature automatically for image retrieval according to mult-dimension feature by the algorithm and study the scope for weight further, and excavate the better features weight, for the popurse of obtaining the results more satisgfied.Finally, we improve the result of image retrieval based on multi-dimension feature. People can not get good effect of multimedia retrieval only by low-level features, for the purpose of solving the problem of semantic gap, and locating the multimedia information by the high-level semantic information people are well known, the paper propose matrix division method to address one high-level semantic framework, with which we can organize the high-level semantic information efficiently. We propose result partial priority algorithm by users retrieval based on matrix division method, by which we adjust branch weight of high-level semantic framework.According to the framework, we give semantic branch structure, with which we involved concepts and basic relation of ontology, make the semantics more instantiation, embodiment. Then the structure is more generality. The semantic branch structure not only enhances the influence of the relationship among concepts on the describing of video key frame and the background knowledge of the video key frame, but also enhances the relationship among video key frame. The three combined semantic modules can describe different types of video key frame semantic concepts. The module in different types contains different semantic concepts which are organized according to the semantics for the type. By the method, we can improve the reusability of the modules and the capacity for semantic data is reduced.The high-level semantic features of the video key frame are addressed in the form of soft vector that can give each attribute of the soft vector different values. Thus we conquer the disadvantages that each attribute of the Boolean vector are the same values. The experiment shows that soft vector method is better than Boolean method in aggregated effect.we design and realize a multimedia retrieval system according to low-level and high-level semantic information. By combinding many technologies, we carry out multimedia analyzing according to multi-dimension feature based on muti-dimension storing structure. The system takes distributed structure and enables users access the system from remote node.At the end of the thesis, we concluded the work, analyze and give the recommendation for further research in the future.