Abstract: A typical geographical map consists of a large number of lines, curves and symbols representing various physical entities and their spatial relationships. Based on the principle of Freeman's chian code, the authors propose a "predicting-while tracing" technology, which is suitable for the contour recognition and representation of the elements of a map. In the process of recognition,, the prior knowledge of maps is used sufficiently. The analytic features of the elements of a map are extracted through recognition with learning. Finally, experiment results on a number of binary valued large scale maps are also presented.