A Fast Image Retrieval Method Based on A Quantization Tree

Abstract

Traditional content-based image retrieval technology expresses the content of images by low-level features, giving rise to the “semantic gap” problem. Therefore, it is of great significance to obtain better accuracy of retrieval by fusing the visual spatial semantic content of images into content-based image retrieval algorithms. To help retrieval algorithms improve the ability of image analysis and understanding and improve the accuracy of the retrieval algorithm, in this chapter, we propose a new image retrieval methodology that begins by partitioning images into small overlapping image patches, on which feature vectors in the form of color histograms are exacted. Next, feature vectors extracted from a small number of training images are clustered to obtain the knowledge base consisting of visual vocabulary words, and images in the whole database are parsed by the knowledge base and their size-reduced versions are stored separately and subsequently used as an index. Finally the query image is partitioned and parsed, its similarities with the indexed images in the database are calculated, and the most similar images are output. The focus of this paper is on a fast visual vocabulary tree (a fast approximate nearest neighbor search tree), Quantization Tree, to quickly parse the query images so as to partially resolve the semantic gap problem between the semantic image content and the low-level image features. Extensive experimental results for object recognition tasks in an indoor environment as well as in an outdoor environment demonstrate the effectiveness and the efficiency of the Quantization Tree.