| In the era of Internet plus,the network has become a common source of information for people's daily life.The multidedia data on the network shows an explosive growth,and people's demand for information is also increasing.However,the traditional single mode of multimedia retrieval has been unable to meet the needs of people.Cross-media retrieval has been proposed in the context of this technical environment and data requirements,and has rapidly become a research hotspot in the field of computer vision and information retrieval.Cross-media retrieval is based on the content-based multimedia retrieval.However,the content-based multimedia retrieval is mostly for single type of multimedia data,does not support the mutual retrieval between a varity of types of multimedia data.But the goal of cross-media retrieval is to deal with various types of multimedia data,to form a comprehensive understanding of multimedia semantics,and to smoothly retrieval differents types of multimedia data.A key issue to be solved in cross-media retrieval is that how to explore the internal relationship between different types of data,and then to calculate the similarity between cross-media data.In this paper,based on Canonical Correlation Analysis,the underlying heterogeneity features are mapped to the new isomorphic subspace,while the correlation between different types of multimedia data is been maximumly reserved.Similarity of different types of multimedia data are comparable in this subspace.On the basis of this,we proposes a method of canonical correlation analysis based on multi kernel function to solve the problem of the dependence of canonical correlation analysis on data linearity and the difficulty of choosing kernel function.In the evaluation of the experimental results,we use precision and recall as the evaluating indicator.Experimental results show that this method can not only achieve good performance,but also the accuracy is higher than other traditional cross-media retrieval algorithms. |
PDF Full Text Request