The perception of jamming types is very important for protecting our radar in complex electromagnetic environments. Radar active deceptive jamming based on digital radio frequency memory (DRFM) has a high coherence with the target echo, which confuses the information of the target echo and achieves the effect of hiding the real target. Traditional deceptive jamming recognition methods need to extract complex features and artificially set classification thresholds, which is inefficient. The existing neural network-based jamming identification methods still follow the pattern of signal modulation-type identification, so there are fewer types of jamming that can be identified, and the identification accuracy is low in the case of low jamming-to-noise ratios (JNR). This paper studies the input of jamming recognition networks and proposes an improved intelligent identification method for chirp radar deceptive jamming. This method fuses three short-time Fourier transform time–frequency graphs disturbed by three consecutive pulse periods into a new graph as the input of the convolutional neural network (CNN). Using a CNN to classify the time–frequency image has realized the recognition of a variety of common deceptive jamming techniques. Similarly, by changing the network input, the original signal is used to replace the echo signal, which improves the accuracy of the jamming recognition in the case of a low JNR.