Abstract: Low probability of intercept (LPI) radar signals are widely used in modern electronic warfare due to their excellent anti-intercept capability. The low peak power of LPI radar signals makes them easily overwhelmed by additive white Gaussian noise (AWGN), which results in low signal-to-noise ratios (SNRs), and poses a great challenge for signal detection and identification. In order to extract the original LPI radar signals from the AWGN background, this paper proposes a deep neural network (DNN) called LPI-U-Net for end-to-end time-domain LPI radar signal enhancement. The network consists of a feature extract module (FEM), a feature focus module (FFM) and a signal recover module (SRM). First the FEM extracts the features of the signal by convolution operation, then the FFM uses convolution and inter-channel attention to further focus on the features that are beneficial to the signal enhancement task, and finally the SRM reconstructs the signal from the features by using the deconvolution operation, thus completing the LPI radar signal enhancement. Simulation experiments show that the performance of LPI-U-Net for LPI radar signal enhancement at low SNR outperforms typical noise reduction methods in conventional signal processing, verifying its feasibility and effectiveness.