The fast detection of communication signals and the exact identification of their modulation types are of importance in practice. Traditional designs use detectors for each modulation type separately thus the computation time would increase as the number of modulation types increases. It is necessary to work out a standard feature vector extraction method to reduce the number of detectors. A novel identification scheme is proposed, with feature vectors being extracted from the time-frequency distribution and identified by an artificial neural network. By adding signal samples of new modulation types and by retraining the neural network, this identification scheme can recognize more modulation types without increase of computation burden. The detail of this feature vector extraction approach is described, the probability of the correct identification of the communication signals in low signal-to-noise conditions is obtained through computer simulations.