Content-based retrieval for European image libraries
IM'99 Proceedings of the 1999 international conference on Challenge of Image Retrieval
| Hi-index | 0.00 |
With the advent of affordable, large-volume media-storage many people and organisations are moving to digital archiving of images. This includes home photograph collections in the form of Photo-CD, photography agencies, newspapers and the World Wide Web. Unfortunately, searching and retrieving images from these databases are not trivial tasks. The human visual perception system is excellent at high-speed analysis of images and the identification of objects within them. Indeed, humans can obtain information from an image far faster than from a textual representation of the scene depicted. However, the main methods of searching an image database are text-based, employing indexing, filenames, subject tags and so forth. Few methods are available which allow pictorial information to form the search-key for a database query. This paper describes a system which has been developed to allow queries on an image database using feature data about regions within each image. The tasks carried out by the system can be divided into two parts: preprocessing and database-query. During the preprocessing stage, images are segmented into meaningful areas with each one being assigned a unique identifier. Feature extraction is carried out to obtain contextual information about the regions including colour, size, shape and texture. The feature sets and region maps are stored in the database along with the images for use during the query phase. At the query stage, users specify regions of interest from an existing image. The feature vectors of these regions are used to place Radial Basis Function (RBF) nodes in n-D space. The aim is to place these nodes in the centre of clusters of feature vectors belonging to the same class as the key region. A set of images from the database according to RBF node activation level of regions they contain. The user applies relevance feedback to the resulting set of images stating which are correct matches. This information is used to carry out progressive refinement on the network and thus, hopefully on the system''s performance. This process of labelling and classification is repeated until the user is satisfied with the results. The paper discusses the techniques employed and demonstrates some results.