A new approach to large-scale image recognition for visual search engines

Image and pattern recognition have become a significant task in recent years when most mobile communication devices have integrated cameras. The bag-of-words (BOW) approach is commonly used in information retrieval algorithms, and although originally developed for text, it has been adapted also for visual search. In spite of the similarity, there are different challenges - less effective weighting algorithms and quantization errors reduce the discrimination power of a single visual word. Moreover, the standard BOW algorithm completely ignores the geometric relationship between visual words. In this paper we introduce an efficient retrieval algorithm that accounts for geometric consistency constraints between two sets of local features that can be integrated with an inverted file. The algorithm significantly improves the initial ranking of the search results, promoting the more suitable candidates to the top of the results list. Application of the new algorithm shows that the proposed method outperforms presently available BOW retrieval algorithms even when followed by full geometric verification. Our conclusion is that the new algorithm and its associated data structure could be instrumental in improving image retrieval tasks.