[Enhancements on Off-line Signature Verification and Recognition]
In the past decades, biometrics research hasalways been the focus of interests for scientists and engineers. (Tax, 2010: 166-169) It is an art of science to use physical and behavioral characteristics to verify or identify a person. (Chen, 2010: 1280-4) Particularly, handwriting is believed to be singular, exclusive, and personal for individuals. Handwriting signature is the most popular identification method socially and legally which has been used widely in the bank check and credit card transactions, document certification. (Franc, 2010: 236- 9) Additionally visual pen-tracking interface has been adopted in many new generation portable computers and personal digital assistants (PDA). (Velez, 2009: 134-138) Thus the computer-aid identity verification based on personal signature has been focus of forensic scientists over this two decades. (Frias-Martinez, 2009: 693-704) Signature verification refers to the specific handwriting processing: the task to compare the questioned signature with the collected signatures of a writer. It requires extracting the exclusive and personal character of writings from the signer, irrespective of the written content. (Lazebnik, 2009: 2169-78) In general signature verification follows two approaches: on-line and off-line. The difference of them lies on that the on-line system handles with multiple one-dimensional temporal signals and the off- line system is constrained on the optical image level instead. (Hsu, 2010: 415-25)
The task of off-line signature verification can be defined as verifying a questioned/ unknown signature when a group of genuine signature from the specific person are available to be referred. Like the most of pattern recognition problems off-line signature verification consists of two phases: training and testing. In the phase of training the user isasked to sign his or her name on the blank paper for several times. Then the designed computer program analyzes the exclusive character of these writings. Mostly some numeric features will be extracted as the prototype with respect to the person. When coming upon a questioned signature its writing style will be analyzed by the similar model and the result will be compared with the constructed prototype of the requested identity. The testing decision of likely authenticity is made consequently. (Gonzalez, 2010: 121-125) Signature verification encounters the dilemma brought by the statement of the task: it is aimed at designing a system to distinguish the genuine signatures from the forgery ones. However the personal writing style can change dramatically over different periods and even under different physical status. (Sabourin, 2009: 976-88) Additionally the over fitting is unavoidable since it is impossible to collect all the genuine samples from the signers and all the forgery samples from impostors. That is to say the distribution of training examples is inconsistent with the true distribution of sample space. Under this consideration people have decomposed the problem into more constrained and trivial ones. Essentially researchers have only claimed their systems fit to the task for eliminating random forgery, simple forgery or skilled forgery which is an approximate categorization for the forgeries. (Gonzalez, 2010: 121-125) The researchers have proposed numerous algorithms on this area even ...