Abstract: Jamming recognition is the premise of radar anti-jamming, but the recognition method based on characteristic parameters is greatly affected by noise. In addition, the feature extraction of parameters only can take place in a certain pulse repetition time, so it is difficult to identify some jamming signals with temporal relationship. However, the idea of using features to identify interference is feasible. On this basis, a jamming identification method was proposed, taking a cascade form to join two convolutional neural networks. Based on the Pseudo Wigner-Ville distribution of the signal, this method was arranged to use the single-period time-frequency image to complete jamming pre-classification and the multi-period composite time-frequency image to complete jamming fine classification, and to recognize eight typical jamming types, especially suitable for the pulling off jamming recognition. The experiment results show that the average recognition accuracy of eight kinds of jamming can reach up to 98% on the data sets generated in this paper.