Interactive search techniques for content-based retrieval from archives of images

Through a little investigation by file types it is possible to easily find that one of the most popular search engines has in its indexes about 10 billion of images. Even considering that this data is probably an underestimate of the real number, however, immediately it gives us an idea of how the images are a key component in human communication. This so exorbitant number puts us in the face of the enormous difficulties encountered when one has to deal with them. Until now, the images have always been accompanied by textual data: description, tags, labels, ... which are used to retrieve them fromthe archives. However it is clear that their increase, occurred in recent years, does not allow this type cataloguing. Furthermore, for its own nature, a manual cataloguing is subjective, partial and without doubt subject to error. To overcome this situation in recent years it has gotten a footing a kind of search based on the intrinsic characteristics of images such as colors and shapes. This information is then converted into numerical vectors, and through their comparison it is possible to find images that have similar characteristics. It is clear that a search, on this level of representation of the images, is far from the user perception that of the images. To allow the interaction between users and retrieval systems and improve the performance, it has been decided to involve the user in the search allowing to him to give a feedback of relevance of the images retrieved so far. In this the kind of image that are interesting for user can be learnt by the system and an improvement in the next iteration can be obtained. These techniques, although studied for many years, still present open issues. High dimensional feature spaces, lack of relevant training images, and feature spaceswith lowdiscriminative capability are just some of the problems encountered. In this thesis these problems will be faced by proposing some innovative solutions both to improve performance obtained by methods proposed in the literature, and to provide to retrieval systems greater generalization capability. Techniques of data fusion, both at the feature space level and at the level of different retrieval techniques, will be presented, showing that the former allow greater discriminative capability while the latter provide more robustness to the system. To overcome the lack of images of training it will be proposed a method to generate synthetic patterns allowing in this way a more balanced learning. Finally, new methods to measure similarity between images and to explore more efficiently the feature space will be proposed. The presented results show that the proposed approaches are indeed helpful in resolving some of the main problems in content based image retrieval.