Signature recognition is probably the oldest biometrical identification method with
high acceptance. Although automated signature verification has been studied for more
than 30 years, present-day offline signature verification systems still can not achieve
better error rates than 10%. Our research aims constructing an efficient off-line
signature analyzer, which can reconstruct the signing method and several hidden features
like velocity or strokes, and use these features by a classifier based on neural network.
In contrast with typical applications, our solution is able to take both global and
local features of the signatures to consideration. Our tests have shown that at some
signatures we can achieve an error rate of 3%. Other ones resulted in error rates
between 20 and 40%. Hence the way to improve the efficiency of our system is to improve
the efficiency of all the phases it is consist of. This paper focuses on classification,
the last phase of signature verification.
