Adaptive Resonance Theory (ART) neural network is analyzed in this paper. A character recognizer is designed and implemented based on the standard ART1 network, aiming at the shortcomings of the standard ART1 network, which concludes the unstablity of network learning and the over sensitiveness to the input sample sequence. This paper gives a method to improve the implementation in C 2 kind of identifier. Experimental validation of these two character recognizer can identify the different character. The improvement method based on standard ART1 network has better stability.