A neural network based method is proposed to classify two dimensional similar figures automatically. As an example of such similar figures, a set of Japanese family crests is employed here. Typical 150 family crests are selected from 7, 500 ones and are treated in the method. Each family crest is digitized by an image scanner and is memorized in a computer as a binary image composed of 96×96 black and white pixels. Firstly, a psychological experiment is executed to investigate similarity among the family crests by 28 testees. Secondly, seventy three kinds of geometrical feature values of a family crest image are defined. Thirdly, a feedforward neural network whose inputs are these features is trained accoding to the results of the psychological experiment. Lastly, classification of unknown family crests is performed by the neural network. It was proved experimentally that the proposed neural network classifies similar family crests successfully.