A novel feature extraction method named blocked winner sequence feature extraction (BWS) method based on SOM is proposed for recognizing large set of Chinese characters. Firstly, a Chinese character sample is blocked to make sub block set according to certain principles; Secondly, every subblock is passed into a SOM network orderly for learning, and the location number of the winner for each sub-block is stored; Finally, location numbers of all winner neurons are combined into a sequence as the feature of the character. Compared with traditional SOM neural networks, the method using orderly combination of multiple winner neurons instead of single winner to represent feature can reduce the network size and computation as well as improve network capacity effectively. Using block can improve the anti-interference ability and the recognition accuracy of the network. The method is used for the extract feature of 3500 Chinese characters in GB2312 Chinese character library with Euclidean distance recognition, and the recognition accuracy reaches 99.897% and 93.274% respectively in the cases of adding 10% and 20% random noise.