The traditional techniques relying on artificial experience for radar emitter feature extraction are cumbersome and have insufficient degree of distinction. To solve the problems, an improved Dual Path Network (DPN) is proposed to automatically extract features and make identification. Firstly the one-dimensional time-domain signal is transformed into two-dimensional time-frequency domain, and then directly input into the DPN for identification. Thus the identification problem of radar emitter is transformed into an image recognition problem. At the same time, considering the problem of feature loss due to too many network layers in DPN, a Triple Path Attention Module (TPAM) is proposed for re-sampling the radar emitter feature map. Then the TPAM is embedded into DPN to form a TPAM-DPN for identifying the radar emitter. Experiments on six common radar signals show that the features extracted by this method are more conducive to improving the radar emitter identification accuracy and are more time-efficient.