Fast Image Retrieval Method Based on Visual Word Tree Word

Fast content-based image retrieval is one of the core problems in multi-media technology and computer vision. Bag-of-words belongs to a currently very popular class of algorithms that work with an efficient visual vocabulary. However, the vocabulary is planar structure to limit the size of vocabulary, representativeness of words and lead to high computational cost. A hierarchical vocabulary scheme called visual word tree is presented. Firstly, features are extracted from training images and hierarchical k-means performed recursively on the descriptor vectors to build a visual word tree, which has k-branch factor and L-layers. The similarity scoring of a database image to the query image is accomplished using inverted files. Due to hierarchical structure, the visual word tree allows a larger and more discriminatory vocabulary to be used efficiently, which has a lower computational cost. We show experimentally a dramatic improvement in retrieval speed on Caltech-101object categories.