Location of address labels on postal parcels using neural networks based on multifeatures

The authors present a method of address-label location on postal parcels using feedforward neural networks. After analysis of about a thousand postal parcels, twenty potentially useful features of address-label candidates were identified. The features were extracted with conventional methods of image processing. Neural networks were trained with these features as inputs to distinguish the address-labels from the set of candidates. Several different neural network architectures were tested. The best one achieved 96% correct classification of the candidates