ImageHawk search engine: Content based image retrieval system

Abstract

This paper describes large-scale content based image retrieval system, Image Hawk search engine. ImageHawk search engine uses 23.4 million images in its gallery. Users have two different methods to make their search: Product Quantization (PQ) and Transductive Support Vector Machine based Hashing using Binary Hierarchical Trees (TSVMH-BHT). Images are first represented with 20480-dimensional Fisher vectors and then binary codes are extracted from Fisher vectors by using these two methods. 256-bit binary codes are used for PQ and 512-bit binary codes are used for TSVMH-BHT. When a query image is given to the search engine, the system returns the most similar 100 images in 30-40 seconds based on the size of the query image. In addition we also describe our new image retrieval dataset created by using ImageCLEF 2013 and report the accuracies of some popular image retrieval methods on this dataset.