Abstarct: The goal of this thesis is to find a new method for on-line signature verification by use of filter bank designed for feature extractions with the help of neural network and support vector machine for classifying .The duty of a signature verification system is accepting a genuine signature and rejecting a forgery signature. Considering that signature verification uses signatures in entrance obtained from screens sensitive to pressure, these kinds of screens extract dynamic features of the signature and its form that makes forging the signature more difficult and as a result , on-line signature verification is more reliable than off-line signature verification. X(t) and Y(t) are extracted from each signature and then previous functions are done on the signature individually. In learning stage, designed filter bank which is especially for signature signals, extracts the important features of signature signal and makes the feature vector of each signature. Among main signatures of each person, 10 are randomly chosen as the reference collection and one signature, which has the least distance with others, is chosen as the template signature. Afterwards, for each signature, the scales test of maximum distance and minimum distance to signature reference collection and also the distance to related model signature is computed. These three scales are normalized and saved in a vector. We add another feature which is the time of signature to these. This vector is manipulated in the used classifier entrance and determines that a signature is genuine or forgery. We use two classifiers of support vector machine and neural network and compare the results. Finally, the correctness rate of suggested algorithms is evaluated for each test signature collection. We use two data baxses, SVC 2004 and Iranian data baxse, for evaluating the suggested algorithm. The correctness rate of suggested algorithm for SVC 2004 data baxse and Iranian data baxse is relatively 97% and 97/2% respectively.