A Study of Language Model for Image Retrieval

Abstract

Recently, various language model approaches have been proposed in the information retrieval realm, with their promising performances in general document and Web page retrieval applications. Based on these achievements, in this paper, we investigate and discuss whether language model approaches can be adapted to content based image retrieval (CBIR), based on the “bag of visual words” image representation. A critical element of language model estimation is smoothing, which adjusts the maximum likelihood estimation to overcome the data sparseness problem. Therefore, we perform extensive studies over different smoothing methods, strategies, and parameters, by showing their impacts to the retrieval performances. Experiments are performed over two popular image retrieval databases, together with some insightful conclusions to facilitate the adaptation of language model approaches to CBIR.