The Research Of Multimedia Retrieval Method Based On Feature Subspace

With the development and popularization of the Internet technology, as well aswidely used of high-capacity storage device and digital equipment, multimedia data,especially the image data increases exponentially. Therefore, the effective managementand application of image database becomes particularly important. As it doesn’t adapt torich low-level visual features of image data, the efficiency of general text search enginefor image data is inefficient. Then, how to search required images from mass imagedatabases accurately and efficiently has become a hotspot in the field of multimediaretrieval research in recent years.Based on traditional content-based image retrieval techniques simultaneously,considering the characteristics that caused by high dimension of visual featurescombination, as well as the following “dimension disaster” problem. This paperconducts a study about the image dimensionality reduction and the semantic subspace.On one hand, local linear regression model is used to calculate the prediction error oftarget subspace. On the other hand, inter-class scatter maximization and intra-classscatter minimization are fused into the objective function as constraints. What’s more, asemi-supervised optimal subspace algorithm is also proposed based on local predictionerror minimization.In addition, for the “semantic gap” between low-level features and high-levelsemantics, this paper integrates method that support vector machine (SVM) learningwhich is based on particle swarm optimization (PSO) to relevance feedback. It not onlymaintains the difference between classes effectively, but also remains stable as a wholeand appropriate individual difference if the individuals are in the same. At the same time,it also improves the effect of the classification.In the evaluation of the experimental results, we use two standards, precision andrecall. The comparative experiment shows that the method proposed by this paper ismore effectiveness and superiority with classic multimedia retrievals in various aspects.