Multimodal indexing based on semantic cohesion for image retrieval

This paper introduces two novel strategies for representing multimodal images with application to multimedia image retrieval. We consider images that are composed of both text and labels: while text describes the image content at a very high semantic level (e.g., making reference to places, dates or events), labels provide a mid-level description of the image (i.e., in terms of the objects that can be seen in the image). Accordingly, the main assumption of this work is that by combining information from text and labels we can develop very effective retrieval methods. We study standard information fusion techniques for combining both sources of information. However, whereas the performance of such techniques is highly competitive, they cannot capture effectively the content of images. Therefore, we propose two novel representations for multimodal images that attempt to exploit the semantic cohesion among terms from different modalities. Such representations are based on distributional term representations widely used in computational linguistics. Under the considered representations the content of an image is modeled by a distribution of co-occurrences over terms or of occurrences over other images, in such a way that the representation can be considered an expansion of the multimodal terms in the image. We report experimental results using the SAIAPR TC12 benchmark on two sets of topics used in ImageCLEF competitions with manually and automatically generated labels. Experi- mental results show that the proposed representations outperform significantly both, stan- dard multimodal techniques and unimodal methods. Results on manually assigned labels provide an upper bound in the retrieval performance that can be obtained, whereas results with automatically generated labels are encouraging. The novel representations are able to capture more effectively the content of multimodal images. We emphasize that although we