Target-Aware Recurrent Attentional Network for Radar HRRP Target Recognition
Introduction
A High-Resolution Range Profile (HRRP) is composed of the amplitude of the coherent summations of the complex returns from target scatterers in each range cell, which represents the projection of the complex returned echoes from the target scattering centers onto the radar Line-Of-Sight (LOS), as shown in Fig. 1. Since it contains abundant discriminative information, such as target size, scatterer distribution, radar HRRP target recognition has received intensive attention from the Radar Automatic Target Recognition (RATR) community [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19].
For the RATR task, feature extraction is a critical step [20]. Liao et al. [21] extract invariant features based on the integration of the bispectra of range profiles. Du et al. [22] further investigate the recognition methods based on high-order spectra features. In [23], the feasibility of classifying aerial targets using the micro-Doppler signatures is studied, where the features are computed in the form of bicoherence estimates and cepstral coefficients. Although those algorithms based on engineered features are effective for recognition, the features are hand-crafted and rely on the personal experiences.
In recent years, deep learning algorithms have been explored for RATR. The autoencoder models have been improved to extract robust features for HRRP [13], which employ the average profile of each HRRP frame as the correction term to establish the cost function under the Mahalanobis distance criterion, and achieve better recognition performance compared with traditional Stacked Auto-Encoders (SAE) [24]. To model the aspect sectors with few HRRP data, Pan et al. [25] utilize the Deep Belief Network (DBN) to learn discriminative features by sharing latent information of HRRP data globally. However, those networks do not explore the temporal dependence between range cells, although they record the structure information of targets.
Several approaches have been proposed to describe the temporal dependence in HRRP. Pan et al. [15] transform the HRRP sample into a sequence and model the structure across range cells by the Hidden Markov Models (HMM) with a transition probability. To model the HRRP sequence, [4] describes the spatial structure across range by the hidden Markov structure and models the temporal dependence between HRRP samples by the time evolution of the transition probabilities. Instead of utilizing the HMM for RATR, it is possible to employ the Recurrent Neural Network (RNN) which exploits larger state-space and richer dynamics compared to HMM [26]. RNN is a class of artificial neural network with connections between hidden units, which allows the model to exhibit dynamic temporal behavior in the data [27]. It achieves state of the art performance on sequential data for different tasks, such as language modeling [28], speech recognition [26], machine translation [29] and so on.
In this paper, we take advantage of the temporal dependence within each HRRP sequence to use RNN with voting strategy to learn its latent representation for recognition task, where each HRRP is divided into multiple overlapping sequential feature as input. Furthermore, we integrate the basic formulation of RNN with an attention mechanism to reduce the influence of the noise areas and automatically catch the discriminative target areas in HRRP, called Target-Aware Recurrent Attentional Network (TARAN). The label for a HRRP sample is decided based on a weighted sum of the output from each hidden state computed by a multilayer perceptron. Therefore, different from the conventional feature selection methods [31], [32], TARAN aims at evaluating the importance of input at each timestep for recognition rather than individual features and the learned corresponding weights are data-dependent rather than fixed for all of data, which may have different lengths for different HRRP samples.
Inspired by [33], besides the HRRP features in time domain, we also try spectrogram features as input, which is a two-dimensional feature providing the variation of frequency domain with time domain.
The remainder of the paper is organized as follows. We formulate the RATR task with RNNs in Section 2. The paper introduces the proposed architecture, i.e., TARAN for radar HRRP target recognition in Section 3. The parameter setting and detailed experimental results based on measured data are made in Section 4, followed by conclusions in Section 5.
Section snippets
Problem formulation
We formulate the problem in a generic manner, and consider classification of an object into one of K distinct classes based on a D-dimensional data measurement of that object. Specifically, given a set of N input signals where X is the training data set containing N input signals, D is the dimensionality of each signal.
Denote where y(n) denotes the 1-of-K representation of the label of x(n), i.e., y(n) ∈ {0, 1}K, if the label of x(n) is k, and
Proposed model
To apply the RNN model, the input signals need to be first converted into sequential features. In this section, we consider two kinds of sequential features i.e. time sequential feature and spectrogram feature.
Experimental results
To show the effectiveness of the proposed TARAN model, we will compare our models to other counterparts on the measured data according to different criterions in the experiments. We will first briefly introduce our measured radar HRRP dataset and the detailed experiment settings. In the experiments, we will discuss the influence of different parameters on our model and give the detailed analysis based on the quantitative recognition performance. To provide qualitative analysis, we will look
Conclusions
In this article, RNNs are analyzed and utilized for HRRP-based RATR to further take advantage of the temporal dependence between range cells. In order to find the discriminative areas in HRRP, we further propose Target-Aware Recurrent Attentional Network (TARAN), which employs the RNN to encode the HRRP sequence into hidden states and weight the hidden state at each timestep to focus on the discriminative regions data-dependently with the help of attention strategy. All the pieces of the TARAN
Acknowledgment
This work is partially supported by the Thousand Young Talent Program of China, NSFC (61771361), 111 Project (B18039), and the National Science Fund for Distinguished Young Scholars of China (61525105).
References (47)
- et al.
Robust statistical recognition and reconstruction scheme based on hierarchical bayesian learning of HRR radar target signal
Expert Syst. Appl.
(2015) - et al.
Radar high-resolution range profiles target recognition based on stable dictionary learning
Iet Radar Sonar Navig.
(2016) - et al.
Radar HRRP target recognition with deep networks
Pattern Recognit.
(2017) Finding structure in time
Cogn. Sci.
(1990)- et al.
Robust speech recognition using the modulation spectrogram
Speech Commun.
(1998) Backpropagation and stochastic gradient descent method
Neurocomputing
(1993)- et al.
Large margin feature weighting method via linear programming
IEEE Trans. Knowl. Data Eng.
(2009) - et al.
A new feature vector using selected bispectra for signal classification with application in radar target recognition
Signal Process. IEEE Trans.
(2001) - et al.
Bayesian spatiotemporal multitask learning for radar HRRP target recognition
IEEE Trans. Signal Process.
(2011) - et al.
Combinatorial discriminant analysis: supervised feature extraction that integrates global and local criteria
(2009)
Identification of ground targets from sequential high-range-resolution radar signatures
IEEE Trans. Aerosp. Electron. Syst.
A noise-robust radar target classification method based on complex probabilistic principal component analysis
General Assembly and Scientific Symposium
Radar HRRP statistical recognition with local factor analysis by automatic bayesian ying-yang harmony learning
IEEE Trans. Signal Process.
Radar HRRP target recognition based on dynamic multi-task hidden markov model
Proceeding of Empirical Methods in Natural Language Processing (EMNLP)
Radar HRRP target recognition based on k-SVD algorithm
IEEE Cie International Conference on Radar
A two-distribution compounded statistical model for radar HRRP target recognition
IEEE Trans. Signal Process.
Multi-task hidden markov model for radar automatic target recognition
IEEE Cie International Conference on Radar
Radar automatic target recognition based on complex high-resolution range profiles
International Conference on Radar
Automatic target recognition in high resolution sar image based on backscattering model
Dianbo Kexue Xuebao/Chin. J. Radio Sci.
Multi-task hidden markov modeling of spectrogram feature from radar high-resolution range profiles
EURASIP J. Adv. Signal Process.
Robust statistical feature based aircraft identification
IEEE Trans. Aerosp. Electron. Syst.
Neural machine translation by jointly learning to align and translate
Int. Conf. Learn. Represent.
Cited by (79)
BPFAaC: A radar intelligent perception algorithm in 6G integrated sensing and communications scenarios
2024, Expert Systems with ApplicationsQuadruplet depth-wise separable fusion convolution neural network for ballistic target recognition with limited samples
2024, Expert Systems with ApplicationsTarget-attentional CNN for Radar Automatic Target Recognition with HRRP
2022, Signal ProcessingCitation Excerpt :Liu et al. [6] exploit positive alpha-stable distribution to characterize the distribution of HRRP data from sea clutter, and then obtain the more precise detection threshold for constant false alarm rate (CFAR)-based ship detection, thus leading to a better detection performance. In [9,15,22,28], different models (including local factor analysis, multi-task factor analysis, class factorized complex variational auto-encoder and target-aware recurrent attentional network) are developed to extract more information that can depict essential characteristics of the target from airplane HRRP data, and ultimately realize the recognition of airplane types, such as the recognition of Yark-42 and Cessna Citation S/II both belonging to jet plane. In this paper, we intend to achieve the recognition of target types with HRRP, referred to as the radar HRRP target recognition for convenience.