The size of survey data is increasing rapidly, and the automatic classification of objects is becoming more important. The classification is traditionally based, e.g., on point-spread function (PSF) fitting. Recently, several different neural network approaches have been introduced for classification. In this paper we use both self-organized map and learning vector quantization based neural networks for star-galaxy separation. Finally, we test a hybrid algorithm using fuzzy classifier and back-propagation neural networks. We show that different methods give relatively similar results. The classification accuracy is good enough for real data analysis, and selection between different methods must be done based on algorithmic complexity and availability of preclassified training sets.