Compared with the fixed frequency signal, the carrier frequency of frequency hopping (FH) signal is controlled by the pseudo-random codes, so it has better concealment and anti-interference. As an important parameter of FH communication, the modulation mode of FH signal can provide powerful support for combat response, such as identification of friend or foe attribute, positioning and jamming guidance, intelligence information extraction, etc. However, there is still a big gap in modulation recognition of FH signals at the domestic and foreign countries. In this paper, a modulation recognition method of FH signal based on time-frequency transform is proposed. The time-frequency images of different modulation types of FH signals are obtained by SPWVD time-frequency transform, and then the time-frequency images are denoised by convolution autoencoder. Finally, the denoised images are sent to convolution neural network for feature extraction and classification recognition. Simulation experiments prove that the proposed method achieves a good classification effect at low signal-to-noise ratios (SNRs), and achieves a recognition rate of 93.67% at -2dB.