Radar HRRP target recognition with deep networks

Feature extraction is the key technique for radar automatic target recognition (RATR) based on high-resolution range profile (HRRP). Traditional feature extraction algorithms usually utilize shallow architectures, which result in the limited capability to characterize HRRP data and restrict the generalization performance for RATR. Aiming at those issues, in this paper deep networks are built up for HRRP target recognition by adopting multi-layered nonlinear networks for feature learning. To learn the stable structure and correlation of targets from unlabeled data, a deep network called Stacked Corrective Autoencoders (SCAE) is further proposed via taking the advantage of the HRRP's properties. As an extension of deep autoencoders, SCAE is stacked by a series of Corrective Autoencoders (CAE) and employs the average profile of each HRRP frame as the correction term. The covariance matrix of each HRRP frame is considered for establishing an effective loss function under the Mahalanobis distance criterion. We use the measured HRRP data to show the effectiveness of our methods. Furthermore, we demonstrate that with the proper optimization procedure, our model is also effective even with a moderately incomplete training set.