Incremental indexing of objects in pictorial databases

Abstract

Object indexing is a challenging task that enables the retrieval of relevant images in pictorial databases. In this paper, we present an incremental indexing approach of picture objects based on clustering of object shapes. A semisupervised fuzzy clustering algorithm is used to group similar objects into a number of clusters by exploiting a-priori knowledge expressed as a set of pre-labeled objects. Each cluster is represented by a prototype that is manually labeled and used to annotate objects. To capture eventual updates that may occur in the pictorial database, the previously discovered prototypes are added as pre-labeled objects to the current shape set before clustering. The proposed incremental approach is evaluated on a benchmark image dataset, which is divided into chunks to simulate the progressive availability of picture objects during time.