Cross-modal Multimedia Information Retrieval With CCA And Adaboost

The early 1890 s, some scholars opened the exploration of multimedia information retrieval. Among them, the content-based multimedia retrieval on the field has become a new hot topic. At the same time, it has also become a highly anticipated direction on the field of computer vision. Multimedia content-based retrieval is mainly applied technologies such as statistical analysis, pattern recognition, knowledge fields of multimedia databases, human-computer interaction, to solve the time-consuming phenomenon and artificial selection on the presence of subjective differences with the traditional keyword-based multimedia retrieval technology.However, this methodology is still insufficient. Traditional content-based multimedia retrieval technology is mainly used to retrieve a single media type, such as image retrieval, text retrieval, video retrieval, audio retrieval. This research will make people turn to cross-modal multimedia retrieval, which is the cross-media retrieval.Currently, there are three main cross-media retrieval method. The first is a cross-media retrieval method based on analysis of fusion. This method integrates different types of feature. The commonly used method including random class method, the weighted average method, Bayesian estimation method; and artificial intelligence class methods include fuzzy logic, neural networks. The second method is based on association mining. In the process of multimedia semantic understanding, fusion analysis methods has difficult to achieve complementarily and enhanced information, so some scholars proposed association mining method, used to find a deeper meaning inside the data. Commonly used methods include cross-reference relationship, connection relationship model and multimedia relationship diagram. The third method is correlation analysis. There still exits a problem that it is difficult to mapping the lower data to the higher semantic with the first and second methods. Meanwhile, some scholars have pointed out that the multimedia objects exists a correlation on each type of content. Using the correlation methods, not only to solve the underlying differences in heterogeneous media content, but also retains the correlation between variables. The disadvantage is that this method too dependent on the underlying characteristics of the data.According to the shortage of the existing technology, this article has carried on the related research and extension of cross-media retrieval and puts forward two kinds of methods: based on the CCA and Adaboost cross modal multimedia retrieval method and multimedia information retrieval based on multiple mapping fusion method, and the two methods was applied to image and text retrieval in order to prove the effectiveness of the methods. The former by using CCA methods to depict the correlation between image and text feature, using Adaboost approach to feedback, repeatedly adjust the correlation. On the basis of the latter from the former, and two fusion methods are put forward. The first method using logistic regression method to make the image and text feature correlation mapping and semantic mapping to the same space; The second approach using the majority vote continue to adjust the correlation between the image and text features. Merge multiple correlation mapping in order to achieve the best state.The effectiveness and efficiency of these approaches are evaluated by the text and inamge retrieval tasks on two public corpora. Experimental results show that the two methods can achieve the desired results, and has improved the accuracy.