Multiple randomized sub-vectors quantization hashing for high-dimensional image feature matching

This paper proposes a new indexing algorithm based on multiple randomized sub-vectors quantization hashing (MRSVQH) for efficient high-dimensional image feature matching. The proposed MRSVQH algorithm quantizes feature vectors according to the L₂ norms of the randomized sub-vectors and hashes feature vectors to their corresponding hash buckets. Such index structures are built for multiple times in order to increase the searching accuracy. On the query stage, the searching process is limited only in the feature vectors that have the same hash value to the query one. Experimental results demonstrate that our MRSVQH algorithm can significantly improve the performance of nearest neighbor searching of image features in both accuracy and efficiency compared to the classic BBF and LSH algorithms, which in turn makes it particularly suitable for image matching and image retrieval.