Modulation recognition is widely used in both military and civil fields. With the complexity of electromagnetic environment, modulation recognition technology is needed to adapt to the situation of large dynamic SNR and multi signal coexistence. In this paper, a modulation recognition algorithm based on deep learning is proposed, which directly extracts the features of IF signals and simplifies the design of artificial features. Two classifiers are used to ensure the correct rate with lower complexity in the case of single signal, while the accuracy rate and noise robustness are higher than traditional methods when multiple signals with time-frequency aliasing coexist.