Abstarct: The task of signature verification system is to accept genuine signature and reject forgery signature. The available systems register signatures with electrical tools. This signatures which are called on-line signatures, are stored as real time signals. This thesis, we present a new algorithm with use of mellin transform and Mel frequency cepstral coefficients for signature recognition. We extract signals x(t) and y(t) From each signature. We utilize mellin transform for preprocessing and Mel Frequency Cepstral Coefficients for feature extraction and make feature vector for each signature. We select from genuine signatures of each person, 10 signatures as reference set randomly and one signature that has minimum distance from another signatures, as template signature. Then we calculate maximum and minimum distance from refrence set and distance from template signature for each test signature. These three values are normalized and store in a feature vector. We apply this vector to classifiers and decide this test signature is genuine or forgery. We compare the results of applying feature vectors to two different classifiers, linear with PCA and neural network. Finally, we evaluate the authenticity accuracy of presented algorithm for each test signature set. We use two databaxse, SVC2004 and Iranian databaxse, for evaluating the presented algorithm in this thesis. Authenticity accuracy of presented algorithm is %94.7 and %96.2 for SVC2004 and Iranian databaxse respectively.