We propose a novel locality sensitive vocabulary coding scheme to extract compact descriptors for low bit rate visual search. We employ Latent Dirichlet Allocation (LDA) to learn the topic vocabularies of lower dimension to generate compact descriptors. To deal with diverse datasets, LDA model is introduced to subdivide a dataset into groups of images with a topic model, where the code word distributions in each group produce coherent statistics from generative learning. Moreover, our empirical study has shown that the original Bag-of-Word (BoW) is sparse, and the occurrences of non-zero words is coherent within a topic. Our proposed topic-wise vocabulary learning yields a more compact yet discriminative codebook to search images. Given a query image, multiple topics are determined, which is fed into the topic vocabularies to generate more compact topical descriptor. Comparison experiments show our topic-wise locality sensitive vocabulary coding produces more compact and discriminative descriptors than the state-of-the-arts. © 2011 IEEE.
|Original language||English (US)|
|Title of host publication||ICICS 2011 - 8th International Conference on Information, Communications and Signal Processing|
|State||Published - Dec 1 2011|