In this paper, we investigate the behavior of neural network classifiers with the negative data, and develop an off-line handwritten numeral string recognition system based on the neural network classifier that uses negative data when estimating parameters. For numeral string recognition, it is attempted to generate all plausible segmentation candidates by character segmentation, which is followed by recognizing the segmentation candidates and finding an optimal segmentation path. In the preliminary experiments for numeral string recognition, the recognition rate of the classifier trained with both positive data and negative data is much higher than the recognition rate of the classifier trained with only positive data. This is because the classifier trained with negative data produces low matching scores for noncharacters, which enables the numeral string recognizer to exclude non-characters from the segmentation alternatives, so it helps the numeral string recognizer to find correct character segmentation paths.
Handwritten numeral string recognition using neural network classifier trained with negative data
2002-01-01
399048 byte
Conference paper
Electronic Resource
English
Handwritten Numeral String Recognition Using Neural Network Classifier Trained with Negative Data
British Library Conference Proceedings | 2002
|A Hybrid Multiple Classifier System of Unconstrained Handwritten Numeral Recognition
British Library Online Contents | 2005
|A hybrid multiple classifier system of unconstrained handwritten numeral recognition
British Library Online Contents | 2007
|Script Independent Handwritten Numeral Recognition
British Library Conference Proceedings | 2006
|