Classification And Multimodal Data Analysis Based On Cross-media Feature Vector

According to the national public security needs, mass media forming large dataneed to be more efficient and agile processing mode, which requires across mediacalculation theory research and exploration, the goal is to make a breakthrough in thecross media data unified representation and modeling method, correlation reasoningand deep mining, comprehensive search and content synthesis and so on the keyissues.In this paper, the research and design is going to focus on the bottomcharacteristics of the media using its cross-media features which is talking aboutcross-media calculation for public security. Then model the data according to thecross media data unified representation and modeling method. So that theCross-media Elementary Generation Model is constructed, to a certain extent, havingsolved these problems we have talked about above. At the same time, we establishedthe mapping mechanism between the local features and the global features in thecross-media data. Also, we established the corresponding data index structure.Talking about cross-media relevance semantic structural consistent, we constructedthe mixed media characteristic model which is CMEGM, so as to meet the structuralproperties of consensus statement. We have done it in this way because the perceptionof the human brain with the outside world has natural “cross media” characteristicsfrom the point of view of cognitive psychology. Also the CMEGM model has solvedthe semantic unit learning and artificial tagging task which is a very hard work. By theway, every single CMEGM model can simulate a cluster of meaningful data set onhigh-level semantic in the multimedia.Thinking that the CMEGM model has high dimensional features, we put forwardthe Adaptive Nearest Neighbor LE algorithm on the manifold learning area. Thealgorithm separates the continuity with the relationship between different types ofdata based on supervised learning. The experimental result shows that this algorithmis performing better than LE and LLE algorithm. In libsvm kit test environment, we choose different k neighbor values so as to find some important outcomes. Then wefind that the accuracy is higher in our algorithm than that of LE and LLE algorithm.In this paper, we construct the cross-media search engine from the idea of thedesign in the classifier. We use MDS/SVD dimensionality reduction algorithm so asto construct the lower dimensional features, then we calculate the Euclidean distanceas the image content similarity matching model. We also have finished the audiocontent template matching process based on the improved DTW speech recognitionalgorithm. Finally, we have made a completion of the image and audio mutualretrieval which has initially completed the retrieval process from one type of media toanother. At the same time, we established the corresponding cross-media index.However, the index is not from the point of view of the level in constructing whichneeds more complicated and efficient algorithms, we will make a deep research onthat further.