The recognition of modulation type is of great significance in non-cooperative communication systems. With the increasingly complex electromagnetic environment, automatic modulation recognition at low signal-to-noise ratio (SNR) has received more attention. Due to the time/frequency uncertainty principle, it is difficult to capture time-domain and frequency-domain features precisely at the same time. To alleviate this difficulty and improve the modulation recognition accuracy for low SNR signals, we present a new modulation recognition method in this letter. We firstly propose the time-frequency reassignment (TFR) technique to convert one-dimensional signals into a time-frequency images (TFIs), and then we modify and adopt the high-accuracy VGG800 deep network to classify TFIs. Under low SNR this method can effectively distinguish seven major modulation types of signals: binary amplitude shift keying (2ASK), quadrate amplitude shift keying (4ASK), binary frequency shift keying (2FSK), quadrate frequency shift keying (4FSK), binary phase-shift keying (2PSK), quadrate phase-shift keying (4PSK) and 16 quadrature amplitude modulation (16QAM). Experimental results quantitatively demonstrate that the classification accuracy of TFR is higher than short-time Fourier transform (STFT) in cases where SNR ≦ -6 dB.