A structural method is proposed for unconstrained handwritten numeral recognition in this paper. The numeral is first smoothed and the skeleton is obtained. A set of feature points are then detected, and the skeleton is decomposed into primitives. A primitive code is defined to record the information of each primitive, and a global code is derived from the primitive codes to describe the topological structure of the skeleton. According to the global codes, all the numerals are classified into 26 subclasses. Two recognition algorithms have been developed based on the primitive codes. In the first algorithm, prototypes and matching rules are designed by hand. This allows a highly abstract matching rule being designed explicitly. Associated with each recognized numeral, a confidence level is also computed. In the second recognition algorithm, neural networks are used for each subclass, where the learning process can be carried out automatically. Good recognition results have been obtained with digit samples extracted from the NIST database. The performance of the recognition algorithms can still be improved if a more advanced thinning algorithm is used.