Efficient Indexing for Mobile Image Retrieval

In this work, we present a novel indexing method, which contains visual phrase quantization, two-dimensional inverted index and Approximate RANSAC (ARANSAC), for mobile image retrieval. First of all, visual phrase quantization is proposed by mapping the SIFT descriptor to two visual words in an order and representing the image by bag of visual phrases. Then, a two-dimensional inverted index is developed, in which the similarity is computed by counting the co-occurrences of visual phrases between query image and database image. Finally, ARANSAC is investigated to re-rank the top-ranked images returned from the above indexing. Compared with the state-of-the-art methods, our proposed method in mobile image retrieval improves the retrieval efficiency while ensuring the retrieval accuracy.