A segmentation-free recognition of two touching numerals using neural network
Soon-Man Choi, Il-Seok Oh · 1999
The recognition of two touching numerals has been tackled by many researchers with the purpose of recognizing the numeric fields in many document forms. The conventional methods are based on a process with two sequential stages, viz. the segmentation of touching numerals and the recognition of the individual numerals. Due to an unlimited number of different overlapping and touching types, the segmentation-based approach has always had a limited success rate. In this paper, we propose a new segmentation-free method using a neural network. In this approach, two touching numerals are regarded as a single pattern from a pattern source with 100 classes. To obtain a training set for the neural network classifier, we synthesize the patterns by moving two isolated numerals in the NIST database horizontally until they touch. For the test set, we manually extract two touching numerals from the numeric string dataset of the NlST database. By using a modular neural network classifier, promising results have been obtained.