A Novel Approach to Collect Training Images from WWW for Image Thesaurus Building

This paper introduces a novel approach to change gathered images from WWW into training images to build an image thesaurus. The requirements for being training images are a large number of images and with highly relevant to a given concept. To fulfil these requirements, a system should be able to collect a large number of relevant images to a given concept from WWW by the proposed criterion of relevance to the concept for each image. Then, the irrelevant images would be filtered out by the modified hierarchical clustering method based on the weighted combination of 5 MPEG-7 visual descriptors and the proposed criterion of relevance to the concept for each cluster. Upon experimental results, the precision of the set of images generated by the proposed method is about 18% higher than that of the set of images generated by other methods