An off-line system for recognition of free written zipcodes

An off-line system for the recognition of free written zipcodes is presented in the paper. The system is built upon five main components. A division module separates the input zip-codes into isolated numerals. A preprocessing module does all the normalizing, noise removal and thinning. Feature division gives several kinds of feature description and extraction. We have two stages of classification, i.e., coarse and detailed. The former uses an inference procedure based on local structure, with fuzzy pre-classifier being set at the front of the recognition system. The latter uses distance matching on a subset of prototypes, the discrete wavelet transform on some local curve segments an a single-layered BP network on the whole binary image, which reduces significantly the number of prototypes to be matched with the input digit sample. Experimental results have shown the system is efficient on zip-codes collected from practical envelopes.