Radar HRRP Target Recognition via Semi-Supervised Multi-Task Deep Network

Feature representation based on the high resolution range profile (HRRP) is the key technology in radar automatic target recognition(RATR). In this paper, we design a deep-u-blind denoising network(DUBDNet) to extract features with high-noise-stability. The fully convolutional DUBDNet is based on autoencoder and employs fusion layers to transfer input features to high dimensional space. Then radar HRRP shift-robust convolutional neural network(RSRNet) is proposed as the classifier. In the experiment, two radar sensors are used to measure HRRP signals of warplanes and civil airplanes. RSRNet performs high robustness to HRRP time-shift sensitivity via testing translation data. This is also a proof that convolutional neural network(CNN) is shift-robust on HRRP target recognition. Trained with noise-to-noise, DUBDNet can achieve blind-denoising in low signal-to-noise ratio(SNR) and significantly improve the correct recognition rate of targets. When the SNR of input HRRP signals is less than 5 dB, DUBDNet can increase the SNR by 10 dB. When the input SNR is -15 dB, output SNR can be increased by 15 dB and the correct recognition rate of targets can be increased by 15%.