TY - CONF VL - 21 AV - none SP - 287 EP - 291 CY - Innsbruck ID - scholars51 UR - https://www.scopus.com/inward/record.uri?eid=2-s2.0-1442302465&partnerID=40&md5=74dddb901c506b51c4993e78796092b6 N1 - cited By 0; Conference of 21st IASTED International Multi-Conference on Applied Informatics ; Conference Date: 10 February 2003 Through 13 February 2003; Conference Code:62395 A1 - Ahmad Fadzil, M.H. A1 - Intan Mastura, A.M. KW - Algorithms; Data acquisition; Image segmentation; Neural networks; Regression analysis KW - Handwritten digit identification; Histogram operation; Mail sorting; Morphological operation KW - Optical character recognition TI - Identification of handwritten digits Y1 - 2003/// SN - 0889863415 N2 - The paper describes the development of an imaging scheme that recognizes and identifies handwritten digits using a combination of image processing and neural network techniques. The scheme is designed to recognise postcodes so that mail-sorting task can be performed with minimum human intervention, reducing labour cost while increasing speed and accuracy. The imaging scheme is a combination of four processes, namely, data acquisition of postcodes, image pre-processing, image segmentation and neural network-based digit identification. Image pre-processing process is essential to obtain a binary image for segmentation. Here, noise reduction, thresholding and greyscale to binary conversion operations are performed. Postcode image samples are then segmented into separated digit regions. The segmentation process involving normalization and morphological operations produces skeleton image of a digit. This is followed by a histogram operation that produces vertical, horizontal, right- and left-diagonal histograms as input to a neural network identification process. A database of 500 handwritten samples of postcodes is used in trials in which samples are used for image processing, training of the neural network and as test data for performance measure. During segmentation process, 90 of all samples are successfully segmented into single digits. In tests, it is found that the recognition rate is 100 for training data (100 samples) and 80 for test data (50 samples). ER -