Visual word generation is a key observation in obtaining the bag-of-visual-words (BOVW)representation for image retrieval: query image features are mapped to their visual words according to the pre-clustered codebook. In this paper

we propose a novel generation approach based on the spatial correlation of visual words. A visual word co-occurrence table is constructed in the first step. Given the known visual words

a new probabilistic predictor is then presented to acce- lerate the generation of their neighboring visual words. We combine the co-occurrence table with the fast library for approximate nearest neighbors (FLANN)

and test it on the Oxford dataset. Comparisons with representative approaches suggest the efficiency and effectiveness of the new scheme.