In the context of information retrieval by visual content, when the user formulates a visual query, his/her query target is rarely represented by a whole image as assumed in the usual paradigm of query by image example. An image should not be considered as an atomic entity since it is generally formed of a composite set of visual parts which express certain semantics.

    A visual information retrieval system should allow the user to explicitly point out the visual target using the various image components. In our work the goal was to investigate methods to define visual search keys which allow the user to express this visual target and to design and efficiently implement these methods.

    The original contributions proposed in this thesis are new approaches which allow the retrieval of images from their various visual components using two distinct query paradigms.

    The first paradigm is the query by region example. It consists in retrieving images containing an image part similar to a query visual part. For this paradigm we have designed an approach of coarse segmentation into regions followed by a fine description of these regions. Coarse regions, extracted by our new unsupervised segmentation algorithm from images in the database, represent visually salient components in each image. This decomposition allows the user to separately point out a region of interest for his/her query. Query by similar regions in the image database relies on a new region descriptor (ADCS). It provides a fine, compact and adaptive characterization of region photometric appearance, in order to take into account the specificity of a database of region descriptors. In this new approach, segmentation is fast and extracted regions are intuitive for the user. Fine description improves the similarity of retrieved regions compared to existing descriptors, thanks to the increased accuracy of region content description.

    Our second contribution concerns the development of a new image query paradigm by logical composition of region categories. This paradigm has the advantage of providing a solution to the "page zero" problem. It allows the user to attain images, if they exist in the database, which are close to the mental representation of the user visual target. No image nor region example is necessary to formulate the query. This paradigm relies on the unsupervised generation of a region photometric thesaurus constituted by the visual summary of regions in the database. To formulate a query the user can access this summary directly by means of logical composition operators on these different visual parts. Note that a visual item in this summary is a representative of a photometric class of regions. Logical queries on image content relate to those in text retrieval. The originality of this paradigm opens rich perspectives for future work in visual information retrieval.

• Related publications
• More information on the first paradigm
• Screenshots illustrating the new paradigm