A Time-Frequency Image Denoising Method via Neural Networks for Radar Waveform Recognition
Zhaocheng Hu, Jie Huang, Dexiu Hu, Zewen Wang
Abstract
Time-frequency images (TFIs) of radar emitter signals can reflect intra-pulse modulation information and be utilized to recognize waveforms, which is helpful for radar emitter identification. However, radar signals are usually interfered with by noise, therefore the robustness of TFIs needs to be improved. This letter proposes a TFI denoising method based on neural networks which can contribute to identifying radar waveforms at the low signal-to-noise ratio (SNR). Firstly, a neural network is trained to generate denoised spectrums from time-domain signals directly. Then the discrete Fourier transform operation in the short-time Fourier transform is replaced by the trained neural network to obtain denoised TFIs. Finally, several advanced convolutional neural networks (CNNs) are employed to classify these TFIs to realize waveform recognition and verify the effect of the proposed denoising method. In the experiment, 16 types of waveforms are simulated for recognition through several methods. The results show that the proposed method reduces the noise in the TFIs significantly, leading to the recognition accuracy being improved under different CNN structures.