Improvement in image retrieval performance of vocabulary tree by adding index storage array and multiple search algorithm

Vocabulary Tree (VT) is one of offline learning-classifier to deal with large number of image set efficiently by combining bag of words concept with tree structure. Bag of words concept makes our classifier possible to return robust classification results regardless of image size, rotation, and other noises. Tree structure can give us very fast classification testing time. But, because of limitation on memory size and tree-construction time, we cannot make very large size VT to get better performance. In this paper, we suggest a multiple search algorithm which returns similar retrieval performance with smaller VT than existing method. In other words, only using small size VT, we can get expected performance because our new algorithm searches tree multiple times. Index storage array can be used when we try to search tree continuously. This array requires very small additional memory, so we can achieve the benefits of memory size with negligible loss for retrieval performance.