Radar signal recognition is an important topic in electronic countermeasures, which faces great challenges in low signal-to-noise ratio (SNR) scenarios. To address this issue, a new method for radar signal recognition based on deep feature fusion with multiple time-frequency images (TFIs) is proposed. Specifically, we first transform time-domain radar signals into TFIs, and then utilize a deep neural network to extract informative features for modulation recognition. To make full use of time-frequency features from various TFIs, we propose a multi-channel fusion model that merges the three TFIs of a radar signal obtained through the short-time Fourier transform (STFT), Choi-Williams distribution (CWD) transform and multi-synchronous squeezing transform (MSST) into a comprehensive TFI. An improved VGG-Net is employed to perform feature extraction from the combined images and modulation classification for radar signal recognition. Experimental results demonstrate that the proposed method can achieve an overall recognition accuracy of over 90% for twelve typical radar signals at an SNR of -6dB, indicating its effectiveness under low SNR conditions.