A Holistic Classification System for Check Amounts Based on Neural Networks with Rejection

A holistic classification system for off-line recognition of legal amounts in checks is described in this paper. The binary images obtained from the cursive words are processed following the human visual system, employing a Hough transform method to extract perceptual features. Images are finally coded into a bidimensional feature map representation. Multilayer perpeptrons are used to classify these feature maps into one of the 32 classes belonging to the CENPARMI database. To select a final classification system, ROC graphs are used to fix the best threshold values of the classifiers to obtain the best tradeoff between accuracy and misclassification.