Recent content-based image retrieval systems offer an interactive visual browsing of images databases. These methods perform a classification of images (offline) into a search tree for users browsing (online). This approach shows three main problems:1) The size of decriptor vector (n>100) makes distance computing sensitive to dimensionality curse,2) Having many different kinds of attributes into descriptor vector does not help classification,3) In general, classification does not take in consideration users' search context. In this work, we propose a method based on building hierarchical signatures having small increasing sizes, this allows to take users' search context into consideration. Our method tries to reproduce human vision behavior. Descriptor vector contains attributes coming from multiresolution analysis of images. These attributes are organized by an expert of the images domain into several hierarchies made of four signature vectors of small but increasing sizes (4, 6, 8 and 10 attributes). These signatures are used to build a fuzzy research tree with k-means classification algorithm (two improvements of this algorithm are given). Online users choose a hierarchy of signature between those built by expert following their search context. A demonstration software has been developed. It uses a dynamic web interface (PHP), optimized image processing tasks using Intel IPP and OpenCV libraries, a MySQL relational database for storage and indexation, a Java3D interface to see images classification results. A testing psycho-visual protocol has been proposed. Results on Trans'Tyfipal paleontological images database are presented and offer good answers to users' queries. Our method gives good results in response time and accuracy during visual browsing.
