Signature Verification Incorporating The Prior Model

Signature has been used widely in the fields of Finance and Security as a valid representative of personal identity. Within the latest 20 years, much work has been done in the automatization of signature verification. However, most of current signature verification systems require all kinds of forgeries to be provided in the training process, which causes much trouble in practice. We propose a new system architecture incorporating a prior model to handle this problem.We implement an effective preprocessing method to extract gray signature traces by using a two-level signature mask, which is obtained through local contrast enhancement, adaptive enthresholding, dilation, and bridge operations.The thesis chooses four kinds of complementary features on two scales, including global features, grid gray feature, texture features and high pressure area feature. We make feature selection based on experimental results. Instead of fix grids, adaptive grids derived from the pixel projection histograms are employed.The proposed system ensembles multiple classifiers based on Boosting algorithm. This integrated classifier works in a similar way to the serial integration mode in the training stage, and to the parallel integration mode in the classification stage. It achieved low error rates by combining merits of two modes.The new system architecture is proposed from the point of practice. It incorporates the prior knowledge and doesn't need simple forgery samples in the training process. Assume the signature feature spaces have the similar distributions for all users. In the training stage, two integrated classifiers are trained separately when the training set contained simple forgery samples or not, and a mapping function is built based on parameters of two classifiers. In the practice stage, an integrated classifier is trained for the current user, whose training set is composed of genuine signatures, and random signatures selected from genuine signatures of existed users registered in the system. Then we build the integrated classifier after applying the mapping function, and make classification using the resultant classifier.In the end, experimental results are given. They show our proposed system is effective in this application environment.