Theoretical study on scalable hashing method for large scale nearest neighbor search

Abstract

Large scale nearest neighbor search is of great importance to many applications such as Information Retrieval and Duplicate detection of document or image. The difficulty of this problem not only comes from the requirement of low computational complexity, but also of unsupervised learning. Recently, many research results show that similarity preserving hashing is a promising approach to this problem. However, most existing hash methods are designed only for similarity preserving property which cannot directly guarantee optimal performance of Receiver Operating Characteristic (ROC) curve. In this paper, we analyze the condition on hash function to achieve optimal False Alarm Rate (FAR) and Detection rate ($P_D$) and propose a novel hashing method based on the condition. The experimental result shows that proposed method outperforms other methods for various applications including duplicate image detection, large-scale image search, and object recognition.