Fundamental in defense, Radar Electronic Warfare (REW) requires adaptation to current threats. Automatic recognition algorithms for intrapulse modulations (ATR) of Low Probability of Interception (LPI) radar signals are essential in REW. Existing LPI signal ATR methods combine the Choi-Williams Distribution (CWD) pre-processing technique with Convolutional Neural Networks (CNN). This work proposes two new ATR combinations, based on SqueezeNet and GoogLeNet CNN. Both used the Smoothed Pseudo-Wigner-Ville distribution (SPWVD) pre-processing technique as an alternative to CWD. Replacing CWD by SPWVD was based on the hypothesis that the latter usually provides higher resolutions than the former. The proposed ATR overcame the SOTA ATR, achieving a 99.06% accuracy, under noisy environments and providing evidence to the hypothesis raised. Experiments involved two datasets with 13 types of modulations and 806,000 samples each.