Siganture verification is one of the oldest methods of biometric identification, which is – due to practicality and effectiveness – still widely used. In the case of artificial intelligence based signature verification, a distinction is made between online and offline scenarios depending on whether the signing and the verification processes coincide, or are separated in time. Off-line signature verification takes has a real practical benefit, because in practice most of the time only a signed paper is available. This method is also much more user-friendly and widely applicable. Even though the field now has a significant background, and people tried to create off-line signature verification system for decades using artificial intelligence in various ways, even today you cannot find one that is capable of crossing the 10 percent margin of error.
In AHR (Signature Verification System) project we tried to build an off-line signature verification system, which decomposes the signature verification process into the smallest possible independent steps and we tried to optimize these steps individually.
This thesis mainly focuses on the classification, the last phase of signature verification. After the review of classification methods and the currently available signature verification systems we discuss two possible solutions for the classification problem. We examine a neural network based and a statistical model based approach. We look at the main challenges of the design, the implementation process, and finally we examine and analyze the results achieved by the classification.