A novel method for radar emitter signal recognition based on time-frequency atom feature is presented. During training, based on the over-complete time-frequency atom dictionary, a few atoms which can separate different kinds of signals best are extracted as a set of fixed feature according to the class separability. During testing, the module of inner product between atoms and signals is used as the input feature for the fuzzy ARTMAP classifier, and the radar emitter signals can be recognized automatically. Experimental results of five kinds of typical radar emitter signals show that this method reduces the computational amount of feature extraction during testing obviously, and the input features have strong concentration within classes and large separability between classes. Our method can achieve high recognition accuracy at the SNR larger than 3 dB.