Unconstrained handwritten numeral recognition using Hausdorff distance and multi-layer neural network classifier
Xuejing Wu, Pengfei Shi · 1999
A method for zip code recognition is presented. 2D binary images are input to HAVNET, a neural network which employs the Hausdorff distance as a similarity metric to train the weights which are required to represent the patterns learned by the network. A new learning rule for HAVNET is also introduced. In this approach, HAVNET is combined with a multilayer neural network. The Hausdorff distance acts as a zip code filter. Only the confusing digits are input into the multilayer network for training or recognition. Experimental results show that the correct-recognition rate with zero rejection is more than 97% for a database used by the Chinese mail sorting system.