Building a navigation structure from a fuzzy relationship for image retrieval

Going one step further feedback querying in integrating user into retrieval process, navigation is the more recent approach to find images in a large image collection using content-based information. However, while properties extracted from images are usually fuzzy data, most of the time a navigation structure will deal with binary links from an image (or a group of images) to another. A trivial solution to get a binary relationship from fuzzy data is to apply a threshold, but this solution not only leads to a loss of information but fails to distinguish noise from interesting elements. In this paper, we propose two techniques to eliminate isolated elements and lead to a structure made of more compact subparts. The first one is based on a variable threshold depending on the number of neighbours. The second one, specific to Galois' lattice, is based on taking into account the existing navigation structure for binarisation of descriptions. Experiments showed that it improves the resulting structure by reducing the number of nodes without loosing information on image description, thus improving user experience.