Abstract: To solve the problem of high memory cost,low retrieval accuracy when the background is changed obviously and reduced efficiency when the size of the database is increased of the BoVW (bag of visual words) method based on Euclidean locality sensitive hashing (E2LSH),a fast retrieval method based on word map chain voting of Hamming embedding was presented on the basis of clustering the feature points through multi-probe locality sensitive hashing．The method constructs a single-table visual dictionary with multi-word mapping and soft assignment,reduces the size of the dictionary to reduce memory consumption．Then a word map chain is constructed with Hamming embedding,and a weighting function is proposed to increase the retrieval accuracy．A weighted voting of the word map chain from matching results is made to accomplish image retrieval．Experimental results show that this method can effectively reduce memory consumption,improve retrieval accuracy,and is applicable to large scale datasets．

Key words: Image retrieval,Multi-probe locality sensitive hashing,Hamming embedding,Word map chain,Bag of visual words