INSTITUTE OF INFORMATION TECHNOLOGIES - BAS

Cybernetics and Information Technologies
Volume 3, No 1. Sofia, 2003, Bulgarian Academy of Sciences


Wavelets' Application for Improving the Neural Network Recognition of Text Images

Dimo Dimov, Sergei Milanov

Institute of Information Technologies, 1113 Sofia, E-mail: dtdim@iinf.bas.bg


Abstract: The paper presents usage of wavelet transform for pre-processing of graphic objects, liable to recognition. It is investigated the possibility for preliminary structuring and reducing of the information, passed to a traditional classifier - nonlinear three-layer perceptron, to improve its recognition opportunities. Experiments performed for recognition of isolated letters are briefly reported. The analysis of the experimental results shows that optimal levels can be defined for reducing and structuring of the input information. A heuristic algorithm is also suggested for a priori evaluation of these levels, i.e. for a preliminary definition of the number of necessary neurons in the input layer of the perceptron.
Keywords: image pre-processing, wavelet transform, nonlinear multi-layer perceptron, structuring and reducing of features for recognition