Narrowband Radar Automatic Target Recognition Based on a Hierarchical Fusing Network With Multidomain Features

Abstract

Narrowband radar automatic target recognition (ATR) is often implemented in a single-feature domain, where recognition performance is limited because only one target property is considered. Fusing features from multiple domains is an effective way to improve recognition accuracy. Existing fusion methods utilize simple fusion strategies, such as concatenation and addition, and the classier design is not considered. In this letter, a hierarchical fusion network (HFN) with multidomain features for narrowband radar ATR is proposed. The HFN contains an intradomain network and an interdomain network with abundant features in multiple domains from the radar echo signal as the input of the network. In the intradomain network, the autoencoder (AE) network is used to learn low-dimensional features and reduce the redundancy of features in the same domain. Moreover, the ratio of within-class distance to between-class distance is introduced into the intradomain network to increase the separability of low-dimensional features. The interdomain network fuses different domain features and obtains the classification result through neural networks. The intradomain and interdomain networks construct a whole framework with a unified loss function, which is jointly optimized via backpropagation. Experimental results on the aerial target data set (drones vs. birds) and the ground target data set (wheeled vehicles, tracked vehicles, and humans) demonstrate that the proposed method is effective for narrowband radar ATR.