Point-wise discriminative auto-encoder with application on robust radar automatic target recognition

Abstract

Deep learning methods have a wide range of applications on various recognition problems for their ability to extract hierarchical features. However, most of the deep algorithms learn features from both the target area and the background area fulfilled with noise or clutter. In radar target recognition applications, the performance of applying deep learning methods directly on data with noise or clutter background is restricted because high-level features can be polluted by a significant amount of irrelevant patterns in the background. Therefore, a deep learning based model, point-wise discriminative auto-encoder, is proposed in this paper to extract noise and clutter robust features from the target area. We bind the original auto-encoder (AE) with a target area extraction net, which can learn the target area automatically from the raw data, to reduce the influence of the noise or clutter background. Moreover, we take advantage of the label information by adding a supervised constraint in our algorithm to help the target area extraction net learn the target area precisely and to extract discriminative high-level features. And then, our proposed method is applied to both high-resolution range profile (HRRP) data and synthetic aperture radar (SAR) images in radar automatic target recognition (RATR) problems. Experimental results on the measured HRRP data and SAR images show the advantages of our proposed method in real applications.