ABSTRACT
A formal description of the proposed retrieval framework based on representative regions is given in this section.
2.1 Extraction of representative regions
In order to find blocks representing different substance in an image precisely, K-means and AP clustering segmentation methods are used. In K-means algorithm, n objects are divided into K clusters and classified depending on the similarity of data. The number of clusters (K) need to be set in advance in K-means; however, the number of objects in different image are not equal, so setting a confirmed value K is inaccurate. K-means is combined with AP algorithm in our method to identify the number of clusters automatically. The strategy we adopt is over clustering and clustering again; categories of objects are usually less than 10 in an image, and accordingly, images are operated by the K-means clustering, where K is set to 10, and then, the results are operated by AP clustering in order to merge the similarity regions to the same category. The reason for not adapting AP along is that the algorithm is complex and time-consuming facing a big database.
