Multimodal image collection summarization using non-negative matrix factorization

The huge amount of biomedical images that are produced every day require of suitable methods to access them in an efficient and effective way. Although there has been an important development in methods to search large information repositories, these methods have been mainly focused on text data, and less work has been devoted to non-text data such as images and video. This paper presents a new method that combines text and visual information in the same latent representation space in which images and text are jointly represented. We also investigate how to select the most representative elements of the collection to build an image collection summary. The proposed method was applied to a collection of histological images and the results where evaluated both qualitatively and quantitatively by an expert. The initial results show that the proposed method is able to build a meaningful summary that can be integrated in an image collection exploration system.