Topic Modeling for Content Based Image Retrieval

Latent Dirichlet allocation (LDA) topic model has taken a center stage in multimedia information retrieval, for example, LDA model was used by several participants in the recent TRECVid evaluation “Search” task. One of the common approaches while using LDA is to train the model on a set of test images and obtain their topic distribution. During retrieval, the likelihood of a query image is computed given the topic distribution of the test images, and the test images with the highest likelihood are returned as the most relevant images. In this paper we propose to project the unseen query images also in the topic space, and then estimate the similarity between a query image and the test images in the semantic topic space. The positive results obtained by the proposed method indicate that the semantic matching in topic space leads to a better performance than conventional likelihood based approach; there is an improvement of 25 % absolute in the number of relevant results extracted by the proposed LDA based system over the conventional likelihood based LDA system. Another not-so-obvious benefit of the proposed approach is a significant reduction in computational cost.