An Experiment with a Neural Network for Handwritten Character Recognition

• This report shows some promising results in using neural networks for handwritten character recognition in real-time. The characters were input from a digitizer and preprocessed to extract different features. These were taken as input to a backpropagation network which was trained to recognize the 29 lower case letters in the Swedish alphabet. Its configuration was 40 input nodes, 30 hidden nodes and 29 output nodes. It was trained with 4 different versions of each character in 15 000 iterations. After that it was able to recognize 100% of two other separately entered versions of the alphabet. The amount of computation needed for classifying a character with this kind of networks puts no restrictions on real-time performance.

Ämnesord

NATURVETENSKAP  -- Data- och informationsvetenskap -- Datavetenskap (hsv//swe)

NATURAL SCIENCES  -- Computer and Information Sciences -- Computer Sciences (hsv//eng)

Nyckelord

Computer science

Datavetenskap

Publikations- och innehållstyp

vet (ämneskategori)

rap (ämneskategori)