| This dissertation presents procedures and results of works on computer signature verification. Two methods were developed and evaluated. First, verification was made using multi-resolution feature representation. This multi-resolution feature representation included global geometric characteristics and wavelet transformations of a signature image. A number of algorithms were developed to extract the global geometric features. A vector quantization classifier and a neural-network classifier were designed to use the multi-resolution representation for verification. Second, verification was made using a grid approach. In this approach, a signature image was divided by a grid and verification was made based on grid features that approximate local structures of a signature image. The grid feature comparison was made using dynamic programming procedures. Results indicated that both systems could detect free-handed forgeries accurately and could also monitor simulated forgeries with reasonable accuracy. |
