Aimed at the problems that traditional emitter signal identification methods often need to carry out artificial feature extraction and signals are difficult to be identified accurately under condition of low SNR environments, a method of emitter signal recognition based on improved UNet3+ network is proposed. By trimming the UNet3+ network hierarchy, the feature fusion ability is retained while the complexity of the network is reduced. The attention mechanism is introduced to optimize the model performance, and a new network model is constructed. The simulation results of eight common radar signals show that the recognition accuracy of the improved model reaches 96.63%. Compared with some classical network models, the total training time is shorter, and the ability to identify the radiation source signal is more effectively under condition of low SNR environments. And the proposed model can also be adapted to the complex electromagnetic environments.