Using Ifs and moments to build a quasi invariant image index

Abstract

Automatic indexing or registration is an essential task for image databases. It allows to archive, organise and retrieve a large amount of images by using inner properties. In this paper, we propose an indexing technique which allows to solve indexing problems due to geometric or photometric transformations, inferred by the different image acquisitions. This approach is based on an invariant partition of the image thanks to the use of interest points (or keypoints) and a characterisation with Ifs parameters or barycentric moments. The research process is based on a similarity measure taking in account a numerical distance and a localisation criterion. This work is based on a local characterisation of the image, we use the interest points to build a triangular partition or a set of triangles. We associate to each polygon a vector containing its photometric properties. In other approaches the keypoints are directly characterised by local invariants. The use of the Ifs parameters to index the image has been studied in early publications, the improvement (robustness against rotations and scaling) comes from the use of the invariant partition and the barycentric coordinates. The research process is particularly important, it uses traditional spatial relations and integrate them with a numerical distance to calculate a score associated to each image.