Automatic on-line signature verification based on multiple models

Automatic on-line signature verification is an intriguing intellectual challenge with many practical applications [I]. This paper proposed a new automatic on-line signature verification system based on three models, a frequency spectrum function model, a shape-related parameter model and a dynamics-related parameter model. Generally, there have been two approaches to automatic on-line signature verification, finction approach ([ 11, [2], [3]) and parameter approach ([4], [ 5 ] , [6] , [7], [SI). The parameter approach rely on comparing specific features of signatures, which are typically global --such as the total time taken, the average or RMS speed. The function approach rely typically on comparing specific functions, such as position coordinates versus time, velocity, acceleration each versus time along the entire signature. The traditional verification schemes are based either on parameter approach or on function approach. But it could be found that the fimction approach and the parameter approach have their own advantages in distinguishing genuine signatures from forgeries respectively. The hnction comparison has the drawback of lumping together the local differences between signature samples irrespective of their causes. In parameter approach, however, we can select features with strong discrimination power from total feature set for each signature. But the different strokes producing sequences of signature trajectory can be obviously reflected in signature hnctions comparison, while can not be reflected in signature global features comparison. This is often the case when a forger did not see the producing process of the signature, only got a visual record of the signature. So this paper proposed a verification scheme based on both the parameter approach and function approach. The entire verification process is split up into two main stages: the function model analysis followed by the parameter models analysis. Only a signature accepted by the function model analysis stage would be had parameter models analysis to decide the authenticity, otherwise, it is rejected as a forgery directly. The aim of the first stage is to reject some forgeries that have different signing style or sequence with the genuine signatures. A kquency fimction model is built in this system by having DIT (Discrete Fourier Transform) on the preprocessed signature coordinates data. The obtained amplitude kquency spectrum functions are compared to their prototypes by the Weighted Cross Correlation. We think the two main aspects of signature data, shape and dynamics, play somewhat complementary roles in distinguishing genuine signatures from forgeries, and it should be avoided that the template of a signature is mainly or completely shape-related or dynamics-related. So we set up two parameter models, a shape-related parameter model and a dynamics-related parameter model, for a signature respectively. Automatic on-line signature verification is feasible only if the system is insensitive to intra-personal variability, but sensitive to inter-personal variability. Although there can be a large number of features available in a signature, not all these features are useful, some even could be harmful for verification. So it is necessary to select personalized features which can overcome the dilemma of intraand inter-personal variability for each signer. A personalized features selection algorithm based on a kind of DP (Discrimination Power) function is used to build parameter models in this system. To find the value of the DP function of a feature, the mean and the deviation of the feature of both a particular signature and the general handwritings are used. And also, a variation of this DP function is applied to be the weight of a feature in parameter model analysis stage. The DP function makes more useful feature be accentuated. These two parameter models are analyzed by fuzzy logic in the system rather than the traditional Euclidean distance metric. It is considered that fuzzy logic is more suitable for the signature verification problem due to the ambiguities of handwritten signatures. Experiments are done to show the effectiveness of the proposed signature verification system.