Y. Suzuki, K. Hatano, and S. Uemura (Japan)
Multimedia Document Retrieval, Integration of Multipul Value
We propose a method to integrate multiple similarity values between each medium in documents and a query. In our multimedia document retrieval system, the document score is calculated by integrating these multiple similarity values, and return a set of documents in descending order of their document scores. An important issue to integrate multiple similarity values are the method of the normalization of each similarity value, and the selection of the mathematical function to integrate the similarity values. To solve first problem, instead of raw similarity values of each medium, we use the “Deviation Value” of each similarity value to normalize the significance of each medium. Moreover, we performed extensive exper iments to find an appropriate mathematical function to integrate the similarity values of each medium into the document scores. As a result, we found that to use the “Deviation Value” is more useful for normalizing multiple similarity values than to use the division of maximum value. We also found that p-norm and one of mathematical functions used at fuzzy set retrieval system were the best mathematical functions in our experiments.
Important Links:
Go Back
