Neural Network-Generated Indexing Features and Retrieval Effectiveness

This article addresses the issue of retrieving images from large archives. It presents a method to deene indexing features describing speciic characteristics of the information contained in the image. Indexing features allow to compute a \degree of similarity" between images. In the method presented here, indexing features are derived from image icons. The latter represent symbolically the image content and are mainly used for browsing. The transition from icons to indexing features is done using a self-organizing map (SOM). In image retrieval systems, SOM-generated indexing features allow to reach high levels of retrieval precision. This is illustrated with ASPECT, a system managing the Zurich archive of solar radio spectrograms. For speciic queries and for recalls less than 10%, a precision above 50% have been reached. It represents about 20% increase compared with a retrieval system based on global indexing features. 1 Overview The eeciency and eeectiveness of a retrieval system for large image archives relies on two main actions: { Browsing through a large number of images allows to visualize roughly but quickly the contents of the archived images. To be quick, browsing uses lossy compressed versions of the images or symbolic image descriptions like image icons (Csillaghy, 1994). { Searching for similar images allows to select the set of images to browse through. This implies the determination of some \degree of sim-ilarity" between images. The degree of similarity relies on indexing features that describe spe-ciic characteristics of the structures contained in the images. Traditional methods to deene indexing features rely, for instance, on text association (Murtagh, 1994), color histograms (Flickner et al., 1995), low-level image properties (Gupta and Jain, 1997) or texture description (Carson et al., 1997). To process a large number of images, the actual values of the indexing features associated with each document must be derived automatically. The au-tomatization is problematic. Traditional methods are usually developed either for conventional photographic pictures (press photographs, museum catalogues etc.) or for small collections of images. Astronomical images archives, on the other hand, are large. Moreover, the archived images are usually noisy and mostly contain diiuse structures. Generally , the methods mentioned above cannot be applied to them. To deene indexing features for astronomical images , another approach must be used. The method presented here uses the information contained in image icons. An icon is composed of a set of boxes describing regions of similar texture (Section 2). Boxes are analyzed …