A Hybrid ANN-Based Technique for Signature Verification

This paper introduces a real-time system for verifying handwritten signatures that relies on a hybrid methodology, for which consistency checking is performed prior to enrolling signatures for further processing. Only the best six signatures are retained out of 10 signatures for each signer, during the enrollment phase, based on the deviation in both the total signing time and the binary pattern of the pen movement. The proposed system consists of three consecutive phases, where the first one is an online approach that is quite similar to the enrollment phase, and acts as an initial bottleneck for the verification process so that simple forgeries are quickly filtered out. The second phase uses a combination of neural networks and linear predictive coding to construct a majority voting committee in a pattern recognition context to decide on the authenticity of the signatures that passed the first phase of the verification process. The third phase is an offline technique that processes the real-time data features after converting them into stationary image frames. A digitizing tablet was used to collect eight features during the implementation of the proposed system resulting in a 2.9% for the false acceptance rate and 8.8% for the false rejection rate. Key-Words: Signature Verification, Pattern Recognition, Hybrid Systems, Neural Networks.