The radar signals of polyphase codes are similar， which is easy to be confused under low signal-to-noise ratio （SNR）. The classic Choi-Williams and other time-frequency distribution methods constrained by the time-frequency resolution are difficult to characterize the details of polyphase codes. Here， we propose an automatic recognition method based on the short-time Fourier transform-based synchrosqueezing transform （STFT-SST） and deep convolutional network. On the feature selection， the STFT-SST is used to radar signals for time-frequency analysis， and a spectrum enhancement algorithm is proposed to enhance the time-frequency features under low signal-to-noise ratio， then the high-resolution feature images are obtained. On the classification network， a nine-layer deep convolution network is designed， and the inception module is introduced to capture the signal’s detailed features. The simulation results show that when the SNR is -8 dB， the average recognition rate for five polyphase codes reaches 91.8%. The recognition performance of the proposed method is better at the low SNR.