To recognize optical character with noise pollution rapidly and accurately, a novel approach for character feature extraction based on statistics and fuzzy membership is proposed. Compared with traditional method, this approach has a higher degree of differentiation in feature space increasing 33% of minimum inter-class distance. Applied in Radical Basis Function (RBF) neural network, under the influence of different font size and image background noise pollution, character recognition rate is up to 75%. Theoretical analysis and experimental results show that, compared with traditional methods, this approach achieves a better anti-noise performance, greater degree of differentiation and lower time and space complexity. It can be simpler, more comprehensive coverage characters' features with wide application. This approach has been applied to the actual system and achieves good results.