Steps towards the development of an automatic classifier for astronomical sources

We present the results obtained in implementing an automatic classifier for astronomical objects. We studied different neural network architectures for the classification of object found in astronomical images (2D case) and we are now implementing a classifier which works both in the image (2D) and time domain. The 2D classifier is based on a Self Organizing Map. The method we describe is adaptive, is trained by examples and doesn't need any training rules. The map is used after training with TAROT objects (Télescope à Action Rapide pour les Objets Transitoires, Rapid Action Telescope for Transient Objects). In this paper, we present the classifiers we tested, and we describe our 2D classifier method as well as the results from simulated and real astronomical images. We present also the next step of classification through our 3D (geometry - time) classifier. In general our method works better than other automatic methods, but needs that an extensive set of all kind of sources, including those rarely encountered, is presented in the training set.