Character Recognition Using Neural Networks EE 368 Semester Project Randy Dimmett Purpose Use a neural network to recognize text in a scanned image Procedure Used Courier New font for sample data and targets Develop network Test with ideal input Test with non-ideal input Procedure Generation of all letters in Courier New 12 pt. 27 inputs each having 108 attributes Procedure Ideal test data Non-Ideal data Tested Neural Networks Linear Associator using Pseudoinverse Rule Up to 9% Accuracy (25% if including spaces) THE ONLY GOOD DAY OF SCHOOL IS THE LAST ONE MLE RBNV FRRM M?Y PN ZBLQPO KZ OJJ LFFU LGN Tested Neural Networks 4-Layers using Back-propagation (2,5,2,and 1 neurons) Reached minimum MSE of .01 Very, Very Bad Results. Tested Neural Networks 5-Layers using Back-propagation (2,5,5,5, and 1 neurons) Reached MSE of about 0. Accuracy less than 6% THE ONLY GOOD DAY OF SCHOOL IS THE LAST ONE YBJ RACS AVST SAZ UZ YEGQRD ZW ZBH GAAZ SBF Troubleshooting Problems with data: Noisy Off-Set Effect Troubleshooting: Adding Noise With noise added to sample data, Linear Associator gives 12% accuracy THE ONLY GOOD DAY OF SCHOOL IS THE LAST ONE XLT OHPZ SIKD CGY BP KPNDCQ OQ ?JM KKLS KNP Troubleshooting: Getting Data Using sample data gotten from scanner, the Linear Associator gives 21% accuracy THE ONLY GOOD DAY OF SCHOOL IS THE LAST ONE QHP ONJV OLOD E?S NP KOEKNP JT SDN LFEO LNG Summary of Results %Accuracy 40 35 30 25 20 15 10 5 0 Character Recognition Accuracy Character Recognition Accuracy incl. Spaces With Ideal Samples With Noisy Samples With Samples from Data Source Conclusions Character Recognition is not a good pattern recognition problem. Results depend greatly on the sample data used. Questions?