Content-based image retrieval in the topic space using SOM and LDA

Content-Based Image Retrieval (CBIR) becomes necessary when dealing with large collections of images. Recently, interesting retrieval methods are based on Bag-of-Words (BoW) model. It allows the projection of an image in a vector space called the word space. Nevertheless, this space can be further refined in order to reflect the semantic content of an image. Indeed, it can be transformed into a lower dimensional vector space called the topic space. In this paper, we propose a retrieval system based on an unsupervised learning: First, a Self-Organizing Map (SOM) is learnt to construct the word space from the feature space. Then, the word space is transformed into the topic space using the Latent Dirichlet Allocation (LDA) model. Once learnt, our model allows to infer a compact and semantic image representation. Experiments are performed using a set of “vehicule” images from Pascal VOC 2007 dataset. Evaluation shows that the proposed retrieval system leads to encouraging results.