Abstract: Aimed at the problems of missing prior information of radio signal, complex operation of manual feature selection and low recognition rate at low SNR, a modulation recognition algorithm based on multi-scale feature residual shrinkage networks (MFRSN) is proposed. The simulation experiment was carried out on the data set containing 11 modulation types, such as PAM4, BPSK, QPSK, 8PSK, QAM16, CPFSK, GFSK, QAM16, QAM64, WBFM, AM-SSB, AM-DSB. The results show that the average recognition accuracy of the signal with low SNR is improved by 2.87% after adding the soft threshold branch, at the same time, multi-scale feature fusion method has better intra class recognition effect compared with other network structures.