Latent Topic Encoding for Content-Based Retrieval

This work presents a new encoding approach based on latent topics which is specially designed to Content-Based Retrieval tasks. The novelty of the proposed Latent Topic Encoding (LTE) lies in two points: (1) defining the visual vocabulary according to the hidden patterns discovered from the local descriptors; and (2) encoding each sample by accumulating the proportion of its local features over topics. Several retrieval simulations using two different databases have been carried out to test the performance of the proposed approach with respect to the standard visual Bag of Words (BoW). Results show that LTE encoding is able to outperform the traditional visual BoW when the retrieval task is performed in the latent topic space.