betaFeedback

WorksEngineering

Hermitian Positive Definite Neural Networks for Matrix-CFAR Geometric Signal Detection in Nonhomogeneous Clutter

Abstract

Matrix information geometry (MIG) detectors exploit the Riemannian geometry of Hermitian positive definite (HPD) matrices to achieve robust object detection in nonhomogeneous clutter. The matrix constant false alarm rate (matrix-CFAR) method is a representative MIG detector that relies on the geometric structures with respect to specific metrics. In most cases, computing representative HPD covariance matrices requires numerical methods, typically involving computationally intensive iterative optimization algorithms. Although explicit formulas are available for certain cases, such as the log-Euclidean (LE) mean, they are generally unavailable for many other matrix means and medians. To address this challenge, we propose integrating the Hermitian positive definite neural network (HPDNet) into the matrix-CFAR framework, replacing conventional numerical optimization procedures with a learned, geometry-preserving mapping that directly estimates representative HPD covariance matrices from data. In the simulations, the framework is evaluated for MIG detectors using both means and medians associated with the affine-invariant Riemannian metric, the LE metric, and the Bures–Wasserstein metric, on both simulated compound-Gaussian clutter and real IPIX sea clutter. In both cases, the HPDNet is trained solely on simulated data to assess its applicability to the study of real data. The results show a significant reduction of computational cost in testing, while maintaining detection performance and robustness.

Introduction

The constant false alarm rate (CFAR) method is an adaptive-threshold detection technique widely used in signal detection [1]. It is designed to distinguish targets from clutter by setting a threshold based on the local noise power estimate. Early CFAR variants operate solely on scalar amplitude or power measurements [2]–[4]. Although they differ in their reference-cell selection strategies, they usually cannot exploit the correlation information contained in multichannel observations, thereby limiting their performance in correlated clutter environments.

Motivated by this drawback, later methods introduced covariance matrix estimation into adaptive detection. Representative covariance-based detectors estimate and exploit the clutter covariance matrix to improve detection performance under homogeneous or nonhomogeneous clutter [5]–[8]. Their performance critically depends on the quality of the sample covariance matrix, which requires a sufficient number of homogeneous training samples [9]. Beyond these estimation-related issues, these methods are formulated within the Euclidean statistical perspective and do not take advantage of the Riemannian geometry of covariance matrices. These observations have motivated the development of matrix-CFAR detectors, in which covariance matrices are modeled as points on Hermitian positive definite (HPD) manifolds, enabling more faithful characterization of clutter statistics [10].

To improve detection performance in nonhomogeneous clutter, various matrix information geometry (MIG) detectors have been developed within the matrix-CFAR framework [11]–[14]. These methods perform detection by measuring geometric distances or divergences between covariance matrices. Since these geometric measures differ substantially in their mathematical formulations, they exhibit different computational complexities and detection characteristics. Thus, the practical deployment of MIG-based detectors is hindered by a trade-off between detection performance and computational efficiency. For example, compared with the log-Euclidean (LE) metric, the affine-invariant Riemannian metric (AIRM) requires computationally intensive matrix operations. This trade-off motivates the development of learning-based approaches that reduce the computational burden of MIG detectors while preserving their detection capability.

To reduce the computational cost, the concept of a neural network has been introduced to matrix-CFAR MIG detectors. Unlike the calculation in traditional geometric measures such as AIRM, the layer computations in neural networks generally require fewer computational resources. As a result, if the layer structures and their connections are appropriate, neural networks would strike a balance between efficiency and effectiveness. In particular, the symmetric positive definite neural network (SPDNet) was proposed to generalize conventional neural network operations from Euclidean space to Riemannian manifolds [15]. The Hermitian positive definite neural network (HPDNet) was later developed as a complex-valued extension of SPDNet, so that it can be directly applied to complex covariance matrices which are commonly encountered in signal processing [16]. These geometric neural networks have recently been explored for covariance matrices in signal detection, showing potential for efficient and robust detection in nonhomogeneous clutter environments [17], [18].

In this paper, we apply the HPDNet to the matrix-CFAR framework to reduce the computational cost of geometric signal detection in nonhomogeneous clutter. The main contributions are summarized as follows:

  1. We propose an HPDNet-based framework for matrix-CFAR MIG detectors for estimating geometric means and medians. Unlike conventional iterative optimization schemes, our framework directly learns the mapping from a set of covariance matrices to their geometric statistics via an HPDNet, eliminating the need for per-test iterative computation and enabling rapid inference once trained.

  1. We apply the proposed method to both simulated compound-Gaussian clutter and real sea clutter. For the simulated data, training and testing are both conducted on compound-Gaussian clutter. For the real sea clutter data, we train the HPDNet solely on simulated compound-Gaussian clutter and test it directly on the real data, thereby investigating the generalization capability of the learned mapping from simulated clutter environments to practical operational scenarios.

  1. We compare the proposed HPDNet-based method with conventional geometric estimators across several Riemannian metrics on both simulated and real sea clutter data. On simulated data, the HPDNet improves the detection ability under the AIRM and LE metric while achieving comparable performance under the Bures–Wasserstein (BW) metric. On real data, the HPDNet achieves superior performance for the AIRM median, LE mean and LE median and maintains comparable performance under the BW metric, although a slight degradation is observed for the AIRM mean. In both simulated and real cases, the detection runtime is substantially reduced except for the LE mean.

The remainder of this paper is organized as follows. In Section II, we briefly review the matrix-CFAR detection formulation and describe the overall detection scheme. In Section III, we introduce the geometric structure of HPD manifolds, review the AIRM, LE, and BW metrics, and propose the architecture of HPDNet, including its layer-wise structure and training strategy. In Section IV, we report the experimental results obtained from both simulated and real sea clutter data, along with a runtime analysis. Robustness analysis for both the traditional and HPDNet-based methods is provided in Section V. Finally, in Section VI, we summarize the main findings of this study and discuss several promising directions for future research.

The Method of Matrix-CFAR

In this section, we describe the objective and basic scheme of the matrix-CFAR method. The aim is to detect whether an object signal is present in the received data returns, under the assumption that the clutter is nonhomogeneous. Let x∈CN\mathbf{x} \in\mathbb{C}^{N} be the received data, where NN denotes the number of observation channels. For simplicity, we assume that the channels are stationary. The detection problem is formulated as a binary hypothesis test between the null hypothesis H0H_{0} (target absent) and the alternative hypothesis H1H_{1} (target present)

{H0:{target is absent,x=c,H1:{target is present,x=ξs+c,(1)\begin{cases} H_{0}: & \begin{cases} \text{target is absent},\\ \mathbf{x}=\mathbf{c}, \end{cases}\\ H_{1}: & \begin{cases} \text{target is present},\\ \mathbf{x}=\xi\mathbf{s}+\mathbf{c}, \end{cases} \tag*{(1)} \end{cases}

where c\mathbf{c} represents clutter in CN\mathbb{C}^{N}, s\mathbf{s} represents the object signal also in CN\mathbb{C}^{N}, and ξ\xi is a complex-valued coefficient.

To exploit the spatial correlation information contained in multichannel observations, matrix-CFAR performs detection on covariance representations rather than raw signal data. A Hermitian Toeplitz covariance matrix is constructed from the estimated autocorrelation sequence as [19]

R=(r^0⋯r^i⋯r^N−1⋮⋱⋱⋱⋮r^i⋱r^0⋱rˉi⋮⋱⋱⋱⋮r^N−1⋯r^i⋯r^0),(2)\mathbf{R}= \begin{pmatrix} \hat{r}_{0} & \cdots& \hat{r}_{i} & \cdots& \hat{r}_{N-1}\\ \vdots& \ddots& \ddots& \ddots& \vdots\\ \hat{r}_{i} & \ddots& \hat{r}_{0} & \ddots& \bar{r}_{i}\\ \vdots& \ddots& \ddots& \ddots& \vdots\\ \hat{r}_{N-1} & \cdots& \hat{r}_{i} & \cdots& \hat{r}_{0} \end{pmatrix}, \tag*{(2)}

where zˉ\bar{z} denotes the complex conjugate of zz, and the autocorrelation coefficient of lag ii is estimated as

r^i=1N∑n=0N−i−1xnxˉn+i,0≤i≤N−1.(3)\hat{r}_{i}=\frac{1}{N}\sum_{n=0}^{N-i-1}x_{n}\bar{x}_{n+i},\qquad0\leq i\leq N-1. \tag*{(3)}

The covariance matrices {R1,R2,…,RK}\{\mathbf{R}_{1},\mathbf{R}_{2},\ldots,\mathbf{R}_{K}\} estimated from all KK units are subsequently processed to generate a representative covariance matrix by

Rg=Φ(R1,R2,…,RK),(4)\mathbf{R}_{g}=\Phi\left(\mathbf{R}_{1},\mathbf{R}_{2},\ldots,\mathbf{R}_{K}\right), \tag*{(4)}

where Φ(⋅)\Phi(\cdot) denotes the geometric mean or median function under a given metric. Consequently, the detection problem (1) becomes the comparison between covariance matrices,

{H0:Rg=RCUT,H1:Rg≠RCUT,(5)\begin{cases} H_{0}:\mathbf{R}_{g}=\mathbf{R}_{\mathrm{CUT}},\\ H_{1}:\mathbf{R}_{g}\ne\mathbf{R}_{\mathrm{CUT}}, \tag*{(5)} \end{cases}

where RCUT\mathbf{R}_{\mathrm{CUT}} is the covariance matrix estimated from the cell under test (CUT). By adopting a threshold γ\gamma, the decision rule satisfies a prescribed false alarm probability. The threshold γ\gamma is estimated by Monte Carlo simulations under the null hypothesis H0H_{0} and the resulting decision rule is

d(Rg,RCUT)⋛H0H1γ.(6)d\left(\mathbf{R}_{g},\mathbf{R}_{\mathrm{CUT}}\right)\mathrel{\underset{H_{1}}{\overset{H_{0}}{\gtreqless}}}\gamma. \tag*{(6)}

Here, d(⋅,⋅)d(\cdot,\cdot) denotes the distance defined by the adopted Riemannian metric. Figure 1 illustrates the scheme of the matrix-CFAR method.

Overview of the matrix-CFAR detection framework

Figure 1. Overview of the matrix-CFAR detection framework.

Riemannian Structures and the Architecture of HPDNets on HPD Manifolds

The proposed framework relies on distances defined on the manifold of HPD matrices. In this section, we will first introduce the geometric tools required throughout the remainder of the paper. In particular, we will review three representative Riemannian metrics, namely, the AIRM, the LE metric, and the BW metric, along with the algorithms for calculating the corresponding means and medians. We then propose the architecture of the HPDNet, followed by the estimation strategy of the means and medians.

Riemannian Metrics

Since covariance matrices are HPD, they can be naturally regarded as points on the HPD manifold, defined as

P(N,C)={P∈GL(N,C)∣P=PH, xHPx>0, ∀x∈CN∖{0}}.(7)\mathcal{P}(N,\mathbb{C}) = \left\{ \mathbf{P} \in GL(N,\mathbb{C}) \mid\mathbf{P} = \mathbf{P}^{\mathrm{H}},\ \mathbf{x}^{\mathrm{H}}\mathbf{P}\mathbf{x} > 0,\ \forall\mathbf{x} \in\mathbb{C}^{N}\setminus\{0\} \right\}. \tag*{(7)}

where GL(N,C)GL(N,\mathbb{C}) denotes the general linear group of all invertible N×NN \times N complex matrices, and (⋅)H(\cdot)^{\mathrm{H}} denotes the conjugate transpose. Equipping a tangent space at every point of an HPD manifold provides a local linear structure for geometric computations. The tangent space at a given point P\mathbf{P} is defined by

TPP(N,C)≅H(N,C),(8)T_{\mathbf{P}}\mathcal{P}(N,\mathbb{C}) \cong\mathcal{H}(N,\mathbb{C}), \tag*{(8)}

where

H(N,C)={X∈CN×N∣XH=X},\mathcal{H}(N,\mathbb{C}) = \left\{ \mathbf{X} \in\mathbb{C}^{N \times N} \mid\mathbf{X}^{\mathrm{H}} = \mathbf{X} \right\},

the vector space of Hermitian matrices, is naturally identified with the local linear space associated with P\mathbf{P}. A Riemannian metric is defined as an inner product for each tangent space that varies smoothly from one point to another, i.e.,

g(P):TPP×TPP→R.(9)g(\mathbf{P}) : T_{\mathbf{P}}\mathcal{P} \times T_{\mathbf{P}}\mathcal{P} \to\mathbb{R}. \tag*{(9)}

Different Riemannian metrics endow the HPD manifold with different notions of distance and geodesic, resulting in different geometric properties and computational costs.

Among various Riemannian metrics defined on the HPD manifold, the AIRM is one of the most widely used. For any tangent vectors X,Y∈TPP(N,C)\mathbf{X}, \mathbf{Y} \in T_{\mathbf{P}}\mathcal{P}(N,\mathbb{C}), the AIRM is defined by [20]

gAIRM(P)(X,Y)=tr⁡(P−1XP−1Y).(10)g_{\mathrm{AIRM}}(\mathbf{P})(\mathbf{X},\mathbf{Y}) = \operatorname{tr}\left(\mathbf{P}^{-1}\mathbf{X}\mathbf{P}^{-1}\mathbf{Y}\right). \tag*{(10)}

The AIRM induces the geodesic connecting two points A,B∈P(N,C)\mathbf{A}, \mathbf{B} \in\mathcal{P}(N,\mathbb{C}), given by

γ(t)=A1/2(A−1/2BA−1/2)tA1/2,t∈[0,1],(11)\gamma(t) = \mathbf{A}^{1/2}\left(\mathbf{A}^{-1/2}\mathbf{B}\mathbf{A}^{-1/2}\right)^{t}\mathbf{A}^{1/2},\quad t \in[0,1], \tag*{(11)}

where (⋅)1/2(\cdot)^{1/2} denotes the principal HPD matrix square root. The length of this geodesic defines the AIRM distance between A\mathbf{A} and B\mathbf{B},

dAIRM(A,B)=∥log⁡(A−1/2BA−1/2)∥F.(12)d_{\mathrm{AIRM}}(\mathbf{A},\mathbf{B}) = \left\lVert\log\left(\mathbf{A}^{-1/2}\mathbf{B}\mathbf{A}^{-1/2}\right)\right\rVert_{\mathrm{F}}. \tag*{(12)}

Here, ∥A∥F=tr⁡(AHA)\lVert\mathbf{A}\rVert_{\mathrm{F}} = \sqrt{\operatorname{tr}(\mathbf{A}^{\mathrm{H}}\mathbf{A})} denotes the Frobenius norm, and log⁡\log denotes the principal matrix logarithm.

The LE metric provides another Riemannian structure on HPD manifolds. The LE metric is constructed by mapping HPD matrices to the vector space of Hermitian matrices through the matrix logarithm. The matrix logarithm log⁡:P(N,C)→H(N,C)\log: \mathcal{P}(N,\mathbb{C}) \to\mathcal{H}(N,\mathbb{C}) is a diffeomorphism onto the vector space of Hermitian matrices. By pulling back the standard Frobenius inner product ⟨A,B⟩=tr⁡(AHB)\langle\mathbf{A},\mathbf{B}\rangle= \operatorname{tr}(\mathbf{A}^{\mathrm{H}}\mathbf{B}) from H(N,C)\mathcal{H}(N,\mathbb{C}) to P(N,C)\mathcal{P}(N,\mathbb{C}), one obtains the LE metric on the tangent space at point P\mathbf{P} [21],

gLE(P)(X,Y)=tr⁡((dlog⁡PX)(dlog⁡PY)),(13)g_{\mathrm{LE}}(\mathbf{P})(\mathbf{X},\mathbf{Y}) = \operatorname{tr}\left((d\log_{\mathbf{P}}\mathbf{X})(d\log_{\mathbf{P}}\mathbf{Y})\right), \tag*{(13)}

where dlog⁡P(⋅)d\log_{\mathbf{P}}(\cdot) denotes the differential of the matrix logarithm. This construction induces a flat Riemannian geometry for HPD manifolds. Under the LE metric, the geodesic between two HPD matrices A\mathbf{A} and B\mathbf{B} is given by

γ(t)=exp⁡((1−t)log⁡A+tlog⁡B),t∈[0,1].(14)\gamma(t) = \exp\left((1-t)\log\mathbf{A}+t\log\mathbf{B}\right),\quad t \in[0,1]. \tag*{(14)}

Accordingly, the induced LE distance is

dLE(A,B)=∥log⁡A−log⁡B∥F.(15)d_{\mathrm{LE}}(\mathbf{A},\mathbf{B}) = \left\lVert\log\mathbf{A}-\log\mathbf{B}\right\rVert_{\mathrm{F}}. \tag*{(15)}

The BW metric originates from optimal transport theory and defines another Riemannian metric for HPD manifolds [22]. Interpreting HPD matrices as covariance matrices of zero-mean Gaussian distributions gives the 2-Wasserstein distance between Gaussian measures, that induces a Riemannian metric for HPD manifolds. This interpretation has made the BW metric increasingly popular in signal processing and machine learning. For any P∈P(N,C)\mathbf{P} \in\mathcal{P}(N,\mathbb{C}) and tangent matrices X,Y∈TPP(N,C)\mathbf{X},\mathbf{Y} \in T_{\mathbf{P}}\mathcal{P}(N,\mathbb{C}), the BW metric is defined by

gBW(P)(X,Y)=12tr⁡(LP(X)Y),(16)g_{\mathrm{BW}}(\mathbf{P})(\mathbf{X},\mathbf{Y}) = \frac{1}{2}\operatorname{tr}\left(\mathbf{L}_{\mathbf{P}}(\mathbf{X})\mathbf{Y}\right), \tag*{(16)}

where LP(X)\mathbf{L}_{\mathbf{P}}(\mathbf{X}) is the unique Hermitian solution to the Lyapunov equation

PLP(X)+LP(X)P=2X.(17)\mathbf{P}\mathbf{L}_{\mathbf{P}}(\mathbf{X})+\mathbf{L}_{\mathbf{P}}(\mathbf{X})\mathbf{P}=2\mathbf{X}. \tag*{(17)}

The corresponding geodesic connecting A,B∈P(N,C)\mathbf{A},\mathbf{B} \in\mathcal{P}(N,\mathbb{C}) takes the form

γ(t)=((1−t)I+tQ)A((1−t)I+tQ),\gamma(t) = \left((1-t)\mathbf{I}+t\mathbf{Q}\right)\mathbf{A}\left((1-t)\mathbf{I}+t\mathbf{Q}\right),
Q=A−1/2(A1/2BA1/2)1/2A−1/2.(18)\mathbf{Q} = \mathbf{A}^{-1/2}\left(\mathbf{A}^{1/2}\mathbf{B}\mathbf{A}^{1/2}\right)^{1/2}\mathbf{A}^{-1/2}. \tag*{(18)}

The corresponding geodesic induces the BW distance

dBW(A,B)=tr⁡(A)+tr⁡(B)−2tr⁡((A1/2BA1/2)1/2).(19)d_{\mathrm{BW}}(\mathbf{A},\mathbf{B}) = \sqrt{\operatorname{tr}(\mathbf{A})+\operatorname{tr}(\mathbf{B})-2\operatorname{tr}\left(\left(\mathbf{A}^{1/2}\mathbf{B}\mathbf{A}^{1/2}\right)^{1/2}\right)}. \tag*{(19)}

Building upon the Riemannian structures, we next address the problem of estimating representative HPD matrices. We introduce the estimation of six representative statistics on HPD manifolds, namely the means and medians under the AIRM, the LE metric and the BW metric. In particular, the LE mean of HPD matrices {R1,R2,…,RK}\{\mathbf{R}_1,\mathbf{R}_2,\ldots,\mathbf{R}_K\} can be calculated by [21]

RLE mean=exp⁡(1K∑k=1Klog⁡Rk).(20)\mathbf{R}_{\mathrm{LE\ mean}}=\exp\left(\frac{1}{K}\sum_{k=1}^{K}\log\mathbf{R}_k\right). \tag*{(20)}

The remaining five statistics do not admit closed forms, and can be obtained by solving optimization problems [20]

Rmean=arg⁡min⁡R∈P(N,C) Gmean(R),(21)\mathbf{R}_{\mathrm{mean}}=\underset{\mathbf{R}\in\mathcal{P}(N,\mathbb{C})}{\arg\min}\ G_{\mathrm{mean}}(\mathbf{R}), \tag*{(21)}

and

Rmedian=arg⁡min⁡R∈P(N,C) Gmedian(R),(22)\mathbf{R}_{\mathrm{median}}=\underset{\mathbf{R}\in\mathcal{P}(N,\mathbb{C})}{\arg\min}\ G_{\mathrm{median}}(\mathbf{R}), \tag*{(22)}

respectively. Here, the objective functions are

Gmean(R)=1K∑k=1Kd2(Rk,R),(23)G_{\mathrm{mean}}(\mathbf{R})=\frac{1}{K}\sum_{k=1}^{K}d^2(\mathbf{R}_k,\mathbf{R}), \tag*{(23)}

and

Gmedian(R)=1K∑k=1Kd(Rk,R).(24)G_{\mathrm{median}}(\mathbf{R})=\frac{1}{K}\sum_{k=1}^{K}d(\mathbf{R}_k,\mathbf{R}). \tag*{(24)}

where d(⋅,⋅)d(\cdot,\cdot) denotes the distance of two HPD matrices, and in the current study, we consider the AIRM (12), LE (15) and BW distances (19).

The specific algorithm depends on the underlying Riemannian metric and the associated optimization problem. For the AIRM mean, the optimization problem is commonly solved via the Riemannian gradient descent method [13]

Rt+1=Rt12exp⁡(1K∑k=1Klog⁡(Rt−12RkRt−12))Rt12.(25)\mathbf{R}_{t+1}=\mathbf{R}_t^{\frac{1}{2}}\exp\left(\frac{1}{K}\sum_{k=1}^{K}\log\left(\mathbf{R}_t^{-\frac{1}{2}}\mathbf{R}_k\mathbf{R}_t^{-\frac{1}{2}}\right)\right)\mathbf{R}_t^{\frac{1}{2}}. \tag*{(25)}

The BW mean can be calculated by the fixed-point algorithm [22]

Rt+1=Rt−12(1K∑k=1K(Rt12RkRt12)12)2Rt−12.(26)\mathbf{R}_{t+1}=\mathbf{R}_t^{-\frac{1}{2}}\left(\frac{1}{K}\sum_{k=1}^{K}\left(\mathbf{R}_t^{\frac{1}{2}}\mathbf{R}_k\mathbf{R}_t^{\frac{1}{2}}\right)^{\frac{1}{2}}\right)^2\mathbf{R}_t^{-\frac{1}{2}}. \tag*{(26)}

For the geometric medians, we will use the Weiszfeld algorithms [23]. For the AIRM median, the update is given by

Rt+1=Rt12exp⁡(Zt)Rt12,(27)\mathbf{R}_{t+1}=\mathbf{R}_t^{\frac{1}{2}}\exp(\mathbf{Z}_t)\mathbf{R}_t^{\frac{1}{2}}, \tag*{(27)}

where

Zt=(∑k=1Klog⁡(Rt−12RkRt−12)∥log⁡(Rt−12RkRt−12)∥F)(∑k=1K1∥log⁡(Rt−12RkRt−12)∥F)−1.(28)\mathbf{Z}_t=\left(\sum_{k=1}^{K}\frac{\log\left(\mathbf{R}_t^{-\frac{1}{2}}\mathbf{R}_k\mathbf{R}_t^{-\frac{1}{2}}\right)}{\left\|\log\left(\mathbf{R}_t^{-\frac{1}{2}}\mathbf{R}_k\mathbf{R}_t^{-\frac{1}{2}}\right)\right\|_F}\right)\left(\sum_{k=1}^{K}\frac{1}{\left\|\log\left(\mathbf{R}_t^{-\frac{1}{2}}\mathbf{R}_k\mathbf{R}_t^{-\frac{1}{2}}\right)\right\|_F}\right)^{-1}. \tag*{(28)}

The update of LE median is given by

Rt+1=exp⁡(∑k=1Kwk(t)log⁡Rk∑k=1Kwk(t)),(29)\mathbf{R}_{t+1}=\exp\left(\frac{\sum_{k=1}^{K}w_k^{(t)}\log\mathbf{R}_k}{\sum_{k=1}^{K}w_k^{(t)}}\right), \tag*{(29)}

where

wk(t)=1∥log⁡Rt−log⁡Rk∥F.(30)w_k^{(t)}=\frac{1}{\left\|\log\mathbf{R}_t-\log\mathbf{R}_k\right\|_F}. \tag*{(30)}

For the BW median, the update is given by

Rt+1=Rt12ZtRt12,(31)\mathbf{R}_{t+1}=\mathbf{R}_t^{\frac{1}{2}}\mathbf{Z}_t\mathbf{R}_t^{\frac{1}{2}}, \tag*{(31)}

where

Zt=(∑k=1K(Rt−12RkRt−12)12dBW(Rt,Rk))12(∑k=1K1dBW(Rt,Rk))−1.(32)\mathbf{Z}_t=\left(\sum_{k=1}^{K}\frac{\left(\mathbf{R}_t^{-\frac{1}{2}}\mathbf{R}_k\mathbf{R}_t^{-\frac{1}{2}}\right)^{\frac{1}{2}}}{d_{\mathrm{BW}}(\mathbf{R}_t,\mathbf{R}_k)}\right)^{\frac{1}{2}}\left(\sum_{k=1}^{K}\frac{1}{d_{\mathrm{BW}}(\mathbf{R}_t,\mathbf{R}_k)}\right)^{-1}. \tag*{(32)}

We remark that several other Riemannian metrics have also been proposed on HPD manifolds (see, e.g., [24], [25]). In this work, we focus on the AIRM, and the LE and BW metrics, as they are among the most widely used metrics in matrix-CFAR detection and statistical learning.

Hermitian Positive Definite Neural Network

The Riemannian metrics define the geometric structures of HPD manifolds. Building upon this geometric framework, HPDNet performs feature transformations directly on HPD matrices while preserving their Hermitian positive definiteness. Its two fundamental layers, the bilinear mapping (BiMap) layer and the eigenvalue rectification (ReEig) layer, are respectively defined as [16]

BiMap⁡W(R)=WRWH,\operatorname{BiMap}_{W}(\mathbf{R})=W\mathbf{R}W^{\mathrm{H}},
ReEig⁡(R;ϵ)=URmax⁡(ϵI,ΣR)URH.(33)\operatorname{ReEig}(\mathbf{R};\epsilon)=U_{\mathbf{R}}\max(\epsilon I,\Sigma_{\mathbf{R}})U_{\mathbf{R}}^{\mathrm{H}}. \tag*{(33)}

where R=URΣRURH\mathbf{R}=U_{\mathbf{R}}\Sigma_{\mathbf{R}}U_{\mathbf{R}}^{\mathrm{H}} denotes the eigenvalue decomposition of an HPD matrix R∈P(N,C)\mathbf{R}\in\mathcal{P}(N,\mathbb{C}) with unitary URU_{\mathbf{R}} and diagonal ΣR\Sigma_{\mathbf{R}}, W∈CM×NW\in\mathbb{C}^{M\times N} is the trainable transformation matrix, MM denotes the output dimension of the BiMap layer, and ϵ\epsilon is a small fixed positive number. In this paper, the HPDNet structure adopted is

Ok=ReEig⁡(BiMap⁡Wk(Rk)),k=1,…,K,\mathbf{O}_k=\operatorname{ReEig}\left(\operatorname{BiMap}_{W_k}(\mathbf{R}_k)\right),\qquad k=1,\ldots,K,
HPDNet⁡(W,{Rk}k=1K)=Avg⁡(O1,…,OK),(34)\operatorname{HPDNet}\left(W,\{\mathbf{R}_k\}_{k=1}^{K}\right)=\operatorname{Avg}(\mathbf{O}_1,\ldots,\mathbf{O}_K), \tag*{(34)}

where Avg⁡\operatorname{Avg} computes the arithmetic mean of the intermediate matrices O1,O2,…,OK\mathbf{O}_1,\mathbf{O}_2,\ldots,\mathbf{O}_K. The architecture of HPDNet with a single layer is illustrated in Figure 2. Although HPDNet can be constructed with multiple layers, this study focuses exclusively on the approximation capability of its single-layer configuration.

Figure 2. HPDNet structure diagram corresponding to (34) with a single layer. The expressions of BiMap and ReEig are given in (33), where W1,W2,…,WKW_1,W_2,\ldots,W_K are trainable parameters. Avg denotes the arithmetic mean.

One important property of this architecture is that it preserves the HPD structure throughout the network. Although the output of the BiMap layer may become rank deficient when WW is not full rank, the subsequent ReEig layer restores positive definiteness by replacing every eigenvalue smaller than ϵ\epsilon with ϵ\epsilon. Consequently, the output of

ReEig⁡(BiMap⁡W(R)),∀R∈P(N,C)\operatorname{ReEig}\left(\operatorname{BiMap}_{W}(\mathbf{R})\right),\qquad\forall\mathbf{R}\in\mathcal{P}(N,\mathbb{C})

is always HPD, allowing the transformation matrices WW to be optimized as unconstrained in CM×N\mathbb{C}^{M\times N}.

To integrate the HPDNet into the matrix-CFAR framework, we adopt only its geometry-preserving BiMap and ReEig layers. Although existing HPDNet architectures further employ a LogEig layer to map HPD matrices into Euclidean vectors [16], such a mapping is unnecessary in our setting because the objective is to estimate a representative HPD covariance matrix rather than a vector representation. Therefore, the LogEig layer is omitted, allowing the network output to remain on the HPD manifold and to be directly used by the subsequent matrix-CFAR detector, where the geometric objectives are directly used as training loss. Accordingly, the loss functions for estimating the intrinsic mean and median are defined using the Riemannian distances as follows:

Lmean(W)=1K∑k=1Kd2(HPDNet⁡(W,{Rk}k=1K),Rk)(35)\mathcal{L}_{\mathrm{mean}}(\mathbf{W})=\frac{1}{K}\sum_{k=1}^{K}d^2\left(\operatorname{HPDNet}\left(\mathbf{W},\{\mathbf{R}_k\}_{k=1}^{K}\right),\mathbf{R}_k\right) \tag*{(35)}
Lmedian(W)=1K∑k=1Kd(HPDNet⁡(W,{Rk}k=1K),Rk)\mathcal{L}_{\mathrm{median}}(\mathbf{W})=\frac{1}{K}\sum_{k=1}^{K}d\left(\operatorname{HPDNet}\left(\mathbf{W},\{\mathbf{R}_k\}_{k=1}^{K}\right),\mathbf{R}_k\right)

The model is trained end-to-end using backpropagation. We employ the AdamW optimizer [26], which decouples weight decay from the gradient update and has demonstrated strong empirical performance for deep neural network optimization.

Detection Performance

In this section, we evaluate the detection performance of the HPDNet framework in nonhomogeneous clutter environments using both simulated compound-Gaussian clutter and real sea clutter data. The performances of the traditional matrix-CFAR and the HPDNet-based approach are analyzed according to their effectiveness and efficiency. All simulations were implemented in PyTorch [27] and conducted on a single GPU.

Simulated Clutter Data

The simulated clutter is modeled as a compound-Gaussian process, which is widely adopted for representing nonhomogeneous clutter due to its ability to capture both local power fluctuations and heavy-tailed amplitude distributions. It therefore provides a realistic benchmark for evaluating matrix-CFAR detectors under nonhomogeneous environments. The compound-Gaussian clutter is expressed as

c=τ⋅u.(36)\mathbf{c}=\sqrt{\tau}\cdot\mathbf{u}. \tag*{(36)}

The texture component τ\tau is assumed to follow a Gamma distribution,

τ∼Gamma⁡(α,β)=βαΓ(α)τα−1e−βτ,τ>0,(37)\tau\sim\operatorname{Gamma}(\alpha,\beta)=\frac{\beta^\alpha}{\Gamma(\alpha)}\tau^{\alpha-1}e^{-\beta\tau},\qquad\tau>0, \tag*{(37)}

where α>0\alpha>0 and β>0\beta>0 denote the shape and rate parameters, respectively, and Γ(⋅)\Gamma(\cdot) is the Gamma function. The speckle component u\mathbf{u} is modeled as a fast-varying complex Gaussian random vector,

u∼CN(0,Σ).(38)\mathbf{u}\sim\mathcal{CN}(\mathbf{0},\mathbf{\Sigma}). \tag*{(38)}

The speckle covariance Σ\mathbf{\Sigma} is chosen as an exponentially correlated Toeplitz matrix, which is widely used to represent clutter returns from multiple pulses. The (i,j)(i,j)-th entry of Σ\mathbf{\Sigma} is given by

Σij=σc2r∣i−j∣ej2πfd,c(i−j)+σn2δij,(39)\Sigma_{ij}=\sigma_c^2r^{\lvert i-j\rvert}e^{j2\pi f_{d,c}(i-j)}+\sigma_n^2\delta_{ij}, \tag*{(39)}

where σc2\sigma_c^2 is the clutter power, r∈(0,1)r\in(0,1) is the one-lag correlation coefficient, fd,cf_{d,c} is the normalized Doppler frequency of the clutter, σn2\sigma_n^2 is the noise power, and δij\delta_{ij} is the Kronecker delta. The jammer signal is added to randomly selected range cells to evaluate its ability to suppress jamming signals,

xi(jammer)=ci+η⋅σc2⋅p(fd,j),1≤i≤N.(40)\mathbf{x}_i^{(\mathrm{jammer})}=\mathbf{c}_i+\sqrt{\eta\cdot\sigma_c^2}\cdot\mathbf{p}(f_{d,j}),\qquad1\leq i\leq N. \tag*{(40)}

where η\eta is the linear jammer-to-noise ratio, σc2\sigma_c^2 is the clutter power, and p(f)\mathbf{p}(f) is the Doppler steering vector defined as

p(f)=1Np[1,ej2πf,ej4πf,…,ej2π(Np−1)f]T,(41)\mathbf{p}(f)=\frac{1}{\sqrt{N_p}}\left[1,e^{j2\pi f},e^{j4\pi f},\ldots,e^{j2\pi(N_p-1)f}\right]^{\mathsf{T}}, \tag*{(41)}

with NpN_p being the number of pulses. The signal model is given by

xN(target)=cN+αt⋅p(fd,t).(42)\mathbf{x}_N^{(\mathrm{target})}=\mathbf{c}_N+\alpha_t\cdot\mathbf{p}(f_{d,t}). \tag*{(42)}

where cN\mathbf{c}_N is the clutter vector of the last range cell, p(fd,t)\mathbf{p}(f_{d,t}) is the target Doppler steering vector defined in (41), and fd,tf_{d,t} is the target normalized Doppler frequency. The target amplitude αt\alpha_t is determined by the signal-to-clutter ratio (SCR):

αt=ρ⋅Pc,(43)\alpha_t = \sqrt{\rho\cdot P_c}, \tag*{(43)}

with ρ\rho being the linear SCR and PcP_c being the estimated clutter power obtained from the median of the first K−1K-1 range cells. By varying ρ\rho, we evaluate the detection probability under different SCR conditions.

In the simulations, we set σc2=10\sigma_c^2 = 10 dB, r=0.9r = 0.9, fd,c=0.1f_{d,c} = 0.1, η=100\eta= 100, and σn2=1\sigma_n^2 = 1 dB. The compound-Gaussian clutter parameters are set as follows: the texture component τ\tau follows a Gamma distribution with shape parameter α=4\alpha= 4 and rate parameter β=3\beta= 3. Two jammers are introduced with a jammer-to-noise ratio of 20 dB and a normalized Doppler frequency of fd,j=0.2f_{d,j} = 0.2. Figure 3 shows an example of the compound-Gaussian clutter power with size 500×500500 \times500. For each metric, 5000 data blocks of size 5×105 \times10 are used to construct the covariance matrix by (2). Thus, the range cell number is K=5K = 5 and pulse number is Np=10N_p = 10. These matrices are fed into the HPDNet as inputs. To determine the matrix-CFAR threshold, 10000 data blocks are used in the Monte Carlo procedure under a constant false alarm rate of Pfa=0.001P_{fa} = 0.001. Then, 2000 data blocks are used to evaluate the detection probability. The target signal injected for detection performance evaluation uses a normalized Doppler fd,t=0.1f_{d,t} = 0.1. The SCR varies from 5 dB to 30 dB in 5 dB increments, where ρ\rho is the corresponding linear SCR value (ρ=10SCR(dB)/10\rho= 10^{\mathrm{SCR(dB)}/10}).

Example of the log power of the compound-Gaussian clutter data

Figure 3. Example of the log power of the compound-Gaussian clutter data.

To establish a consistent stopping criterion for comparing the computational cost of different estimation methods, traditional iterative algorithms terminate when ∥Rt+1−Rt∥F<10−6\lVert R_{t+1} - R_t \rVert_F < 10^{-6}. Unlike the traditional iterative processes that adopt a tolerance, we fix 5000 epochs as the termination condition for the HPDNet. This is because the optimized points of loss functions under different metrics are different. Therefore, in practice, we terminate the training once the training loss converges. As shown in Figure 4, the epoch number of 5000 is sufficient for all of the statistics in this paper to satisfy the condition.

Training loss curves smoothed with a moving window of size 50

Figure 4. Training loss curves smoothed with a moving window of size 50.

The detection results for the compound-Gaussian clutter are shown in Figure 5a, Figure 5c, and Figure 5e. Under AIRM and LE metric, the detection performance of HPDNet based method is superior than that of the traditional matrix-CFAR method, whereas under BW metric, comparable detection performance is maintained. In addition, the HPDNet significantly reduces the detection runtime in all cases, except for the LE mean which admits an analytic formula, as illustrated in Figure 6.

Detection results for AIRM, LE, and BW on simulated and real data

Figure 5. The detection result of simulated data (left) and the real data (right).

Runtime in simulated data

Figure 6. Runtime in simulated data.

Real Clutter Data

We next apply the HPDNet to a real dataset to assess its practical utility. The dataset used here is the IPIX radar sea clutter dataset “19980205_192053_antstep.cdf” from McMaster University [28]. Detailed descriptions of the dataset acquisition and experimental configuration can be found in [29]. The log power of the data is illustrated in Figure 7. The dataset consists of 27 channels and 60000 time samples. Channels 1 and 27 are excluded from the analysis, and the remaining 25 channels are reshaped into 5×105 \times10 data blocks. As a result, the data are divided into 30000 separate data blocks. As in the simulated-data experiments, 10000 data blocks are used to estimate the threshold through a Monte Carlo procedure with a constant false alarm rate of Pfa=0.01P_{fa} = 0.01. Artificial targets with normalized Doppler frequency fd,t=0.2f_{d,t} = 0.2 are added to the last channel of 2000 data blocks to calculate the detection probability PdP_d. The other parameters are the same as those used in simulated data detection. Note that the HPDNet is trained on simulated compound-Gaussian clutter, as in the simulated-data experiments, and is then directly applied to the IPIX dataset without retraining on real data. This setting is adopted to evaluate the generalization capability of the proposed HPDNet-based matrix-CFAR detector across different clutter environments.

Log power of the real clutter data “19980205_192053_antstep.cdf”.

Figure 7. Log power of the real clutter data “19980205_192053_antstep.cdf”.

The detection results for the real sea clutter data are shown in Figure 5b, Figure 5d and Figure 5f. For AIRM median, the HPDNet achieves superior detection performance while reducing computational cost. However, the case of the AIRM mean does not show a comparable result in detection. Under the LE metric, the HPDNet shows better performance in both the mean and median cases. For BW metric, it maintains comparable performance in detection. Figure 8 illustrates that, similar to the simulated case, the testing runtime of the HPDNet is substantially shorter than that of the traditional iterative methods in all cases, except for the LE mean, which can be computed analytically.

Runtime in real data

Figure 8. Runtime in real data.

Robustness Analysis

During the practical detection scenarios, outliers might appear to affect the detection when estimating the means or medians [11, 20]. In this section, we introduce the influence function to evaluate the robustness of HPDNet-based detectors against outliers.

Assume there are KK covariance HPD matrices {Ri}i=1K\{\mathbf{R}_i\}_{i=1}^{K}, and nn outliers {Pj}j=1n\{\mathbf{P}_j\}_{j=1}^{n}. Let Rˉ\bar{\mathbf{R}} be the mean or median of the KK HPD matrices without outliers, and R^\hat{\mathbf{R}} be the mean or median of the K+nK+n HPD matrices including outliers. The mean or median R^\hat{\mathbf{R}} is defined by the following objective functions respectively:

G(R):=(1−ε)1K∑i=1Kd2(R,Ri)+ε1n∑j=1nd2(R,Pj),(44)G(\mathbf{R}) := (1-\varepsilon)\frac{1}{K}\sum_{i=1}^{K}d^{2}(\mathbf{R},\mathbf{R}_i)+\varepsilon\frac{1}{n}\sum_{j=1}^{n}d^{2}(\mathbf{R},\mathbf{P}_j), \tag*{(44)}

or

G(R):=(1−ε)1K∑i=1Kd(R,Ri)+ε1n∑j=1nd(R,Pj).(45)G(\mathbf{R}) := (1-\varepsilon)\frac{1}{K}\sum_{i=1}^{K}d(\mathbf{R},\mathbf{R}_{i})+\varepsilon\frac{1}{n}\sum_{j=1}^{n}d(\mathbf{R},\mathbf{P}_{j}). \tag*{(45)}

Here, ε\varepsilon is a small positive number. Treat R^\widehat{\mathbf{R}} as a perturbation of R‾\overline{\mathbf{R}}, we have

R^=R‾+εH({Ri}i=1K,{Pj}j=1n)+O(ε2),(46)\widehat{\mathbf{R}}=\overline{\mathbf{R}}+\varepsilon\mathbf{H}\left(\{\mathbf{R}_{i}\}_{i=1}^{K},\{\mathbf{P}_{j}\}_{j=1}^{n}\right)+O(\varepsilon^{2}), \tag*{(46)}

where H({Ri}i=1K,{Pj}j=1n)\mathbf{H}\left(\{\mathbf{R}_{i}\}_{i=1}^{K},\{\mathbf{P}_{j}\}_{j=1}^{n}\right) is a Hermitian perturbation matrix. Rewriting (45) and taking the Frobenius norm yield

∥R^−R‾∥F=ε∥H({Ri}i=1K,{Pj}j=1n)∥F+O(ε2).(47)\left\|\widehat{\mathbf{R}}-\overline{\mathbf{R}}\right\|_{F}=\varepsilon\left\|\mathbf{H}\left(\{\mathbf{R}_{i}\}_{i=1}^{K},\{\mathbf{P}_{j}\}_{j=1}^{n}\right)\right\|_{F}+O(\varepsilon^{2}). \tag*{(47)}

Then ∥H({Ri}i=1K,{Pj}j=1n)∥F\left\|\mathbf{H}\left(\{\mathbf{R}_{i}\}_{i=1}^{K},\{\mathbf{P}_{j}\}_{j=1}^{n}\right)\right\|_{F} can be numerically calculated based on the fact that

∥H({Ri}i=1K,{Pj}j=1n)∥F=lim⁡ε→0+∥R^−R‾∥Fε.(48)\left\|\mathbf{H}\left(\{\mathbf{R}_{i}\}_{i=1}^{K},\{\mathbf{P}_{j}\}_{j=1}^{n}\right)\right\|_{F}=\lim_{\varepsilon\to0^{+}}\frac{\left\|\widehat{\mathbf{R}}-\overline{\mathbf{R}}\right\|_{F}}{\varepsilon}. \tag*{(48)}

Then the influence function is defined by [20]

f({Ri}i=1K,{Rj}j=1n):=∥H({Ri}i=1K,{Pj}j=1n)∥F∥R‾∥F.(49)f\left(\{\mathbf{R}_{i}\}_{i=1}^{K},\{\mathbf{R}_{j}\}_{j=1}^{n}\right):=\frac{\left\|\mathbf{H}\left(\{\mathbf{R}_{i}\}_{i=1}^{K},\{\mathbf{P}_{j}\}_{j=1}^{n}\right)\right\|_{F}}{\left\|\overline{\mathbf{R}}\right\|_{F}}. \tag*{(49)}

Note that an analytic method for calculating H({Ri}i=1K,{Pj}j=1n)\mathbf{H}\left(\{\mathbf{R}_{i}\}_{i=1}^{K},\{\mathbf{P}_{j}\}_{j=1}^{n}\right) and hence the influence function can be found in [13].

When outliers are introduced, the number of inputs is increased in the HPDNet, and hence two HPDNets have to be trained, one for determining R‾\overline{\mathbf{R}}, and the other for R^\widehat{\mathbf{R}}. Let W1,…,WK\mathbf{W}_{1},\ldots,\mathbf{W}_{K} be the trained weights in the first HPDNet with KK number of inputs that determines R‾\overline{\mathbf{R}}. In the second HPDNet with K+nK+n inputs for R^\widehat{\mathbf{R}}, the first KK weights are fixed as the ones trained from the first KK input HPDNet, and we only train the remaining weights WK+1,…,WK+n\mathbf{W}_{K+1},\ldots,\mathbf{W}_{K+n} using the objective functions (43) or (44) as the loss function. Consequently, R^∣ε=0=R‾\widehat{\mathbf{R}}|_{\varepsilon=0}=\overline{\mathbf{R}}. In other words, in the absence of outliers, R^\widehat{\mathbf{R}} and R‾\overline{\mathbf{R}} should coincide when computed from the same set of HPD matrices {Ri}i=1K\{\mathbf{R}_{i}\}_{i=1}^{K}.

In the experiments, we adopt the compound-Gaussian distribution (35) to obtain the influence functions. The number of range cells is set as K=50K=50, and the number of pulses Np=5N_{p}=5. The outlier-clutter-ratio is fixed at 40 dB. The other parameters of clutter, such as normalized Doppler frequency for clutter, number of jammers, etc., are the same as introduced in Section IV-A. The number of outliers nn varies from 1 to 40.

As shown in Figure 9, the influence functions of the HPDNet-based estimators exhibit behavior comparable to that of traditional estimators across all three Riemannian metrics, despite some minor differences, particularly in the mean estimators. These results indicate that the proposed HPDNet framework does not compromise robustness against outliers, while maintaining comparable detection performance and significantly reducing computational costs.

Influence functions in traditional and HPDNet-based estimations

Figure 9. Influence functions in traditional and HPDNet-based estimations.

Conclusion

We have proposed an HPDNet-based framework for matrix-CFAR geometric signal detection in nonhomogeneous clutter, in which the iterative estimation of geometric means and medians on HPD manifolds is replaced by a learned end-to-end mapping. This approach eliminates the need for iterative optimization at each test instance, thereby enabling rapid inference once the network has been trained.

Experimental results demonstrate effectiveness of HPDNet across all three Riemannian metrics considered in this study. For simulated compound-Gaussian clutter, HPDNet improves detection performance under the AIRM and the LE metric while achieving performance comparable to that of conventional detectors under the BW metric. For real IPIX sea clutter data, the HPDNet continues to improve detection performance compared to the classical MIG detectors under the AIRM median and both the LE mean and median. Although no further improvement is observed under the BW metric, HPDNet maintains performance comparable to classical BW detectors without degradation.

In terms of computational efficiency, HPDNet substantially reduces computational costs compared with the corresponding iterative algorithms used in MIG detectors, with the exception of the LE mean, which admits an analytical expression. Moreover, since HPDNet is trained solely on simulated clutter and can be applied directly to real sea clutter without retraining, these results demonstrate the generalization capability of the learned mapping. Finally, the robustness analysis confirms that HPDNet preserves the robustness of conventional geometric estimators against outliers.

These findings suggest that learning-based geometric processing on HPD manifolds offers a favorable trade-off between detection performance and computational efficiency, making it a promising approach for real-time signal detection. Although this study focuses on the AIRM, and the LE and BW metrics as representative choices, the underlying idea of replacing iterative optimization with a structure-preserving neural mapping is not limited to these three metrics and could be extended to other geometric measures on HPD manifolds [24, 25].

Another promising direction is to explore alternative Riemannian neural network architectures. Several architectures have been developed for SPD manifolds, including RResNet [30, 31] and SPD Transformer [32, 33], with applications to electroencephalogram (EEG) analysis, facial expression recognition, etc. Investigating how these architectures can be integrated into the matrix-CFAR framework may provide further opportunities to improve detection performance, computational efficiency, and robustness in practical signal detection applications.

Acknowledgments

Acknowledgments. LP is partially supported by JSPS KAKENHI (JP24K06852) and JST CREST (JPMJCR24Q5).

References

  1. [1]R. D. Palmer, “Fundamentals of radar signal processing,” Bulletin of the American Meteorological Society, vol. 89, no. 7, pp. 1037–1039, 2008.
  2. [2]A. Maali, A. Mesloub, M. Djedjou, H. Mimoun, G. Baudoin, and A. Ouldali, “Adaptive CA-CFAR threshold for non-coherent Ir-UWB energy detector receivers,” IEEE Communications Letters, vol. 13, no. 12, pp. 959–961, 2009.DOI
  3. [4]M. Baadeche and F. Soltani, “Performance analysis of mean level constant false alarm rate detectors with binary integration in Weibull background,” IET Radar, Sonar & Navigation, vol. 9, no. 3, pp. 233–240, 2015.DOI
  4. [5]I. S. Reed, J. D. Mallett, and L. E. Brennan, “Rapid convergence rate in adaptive arrays,” IEEE Transactions on Aerospace and Electronic Systems, vol. 10, no. 6, pp. 853–863, 1974.DOI
  5. [8]F. C. Robey, D. R. Fuhrmann, E. J. Kelly, and R. Nitzberg, “A CFAR adaptive matched filter detector,” IEEE Transactions on Aerospace and Electronic Systems, vol. 28, no. 1, pp. 208–216, 1992.DOI
  6. [9]M. Rangaswamy, B. Himed, and J. Michels, “Performance analysis of the nonhomogeneity detector for STAP applications,” in Proceedings of the 2001 IEEE Radar Conference. IEEE, 2001, pp. 193–197.
  7. [10]X. Hua, Y. Cheng, H. Wang, Y. Qin, and Y. Li, “Geometric means and medians with applications to target detection,” IET Signal Processing, vol. 11, no. 6, pp. 711–720, 2017.DOI
  8. [11]X. Hua, Y. Ono, L. Peng, Y. Cheng, and H. Wang, “Target detection within nonhomogeneous clutter via total Bregman divergence-based matrix information geometry detectors,” IEEE Transactions on Signal Processing, vol. 69, pp. 4326–4340, 2021.DOI
  9. [13]Y. Ono and L. Peng, “Towards a median signal detector through the total Bregman divergence and its robustness analysis,” Signal Processing, vol. 201, p. 108728, 2022.DOI
  10. [14]X. Hua, L. Peng, W. Liu, Y. Cheng, H. Wang, H. Sun, and Z. Wang, “LDA-MIG detectors for maritime targets in nonhomogeneous sea clutter,” IEEE Geoscience and Remote Sensing, vol. 61, p. 5101815, 2023.DOI
  11. [15]Z. Huang and L. Van Gool, “A Riemannian network for SPD matrix learning,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 31, no. 1, pp. 2036–2042, 2017.arxiv.org/abs/1608.04233
  12. [16]D. Brooks, O. Schwander, F. Barbaresco, J.-Y. Schneider, and M. Cord, “A Hermitian positive definite neural network for micro-Doppler complex covariance processing,” in International Radar Conference. IEEE, 2019, pp. 1–6.DOI
  13. [17]R. Wang, S. Jin, Z. Cai, Z. Chen, X.-J. Wu, and J. Kittler, “Learning a better SPD network for signal classification: A Riemannian batch normalization method,” IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 11, pp. 19574–19588, 2025.DOI
  14. [18]R. Wang, X.-J. Wu, Z. Chen, T. Xu, J. Kittler, I. Sato, J. Gall, L. Wang, T.-J. Chin, and R. Chellappa, “DreamNet: A deep Riemannian manifold network for SPD matrix learning,” in Computer Vision – ACCV 2022. Springer, 2023, vol. 13846, pp. 646–663.DOI
  15. [19]N. Kang, Z. Shang, and Q. Du, “Knowledge-aided structured covariance matrix estimator applied for radar sensor signal detection,” Sensors, vol. 19, no. 3, 2019.DOI
  16. [20]Y. Ono and L. Peng, “The comparison of Riemannian geometric matrix-CFAR signal detectors,” IEEE Transactions on Aerospace and Electronic Systems, vol. 60, no. 2, pp. 1679–1691, 2024.DOI
  17. [21]V. Arsigny, P. Fillard, X. Pennec, and N. Ayache, “Log-Euclidean metrics for fast and simple calculus on diffusion tensors,” Magnetic Resonance in Medicine, vol. 56, no. 2, pp. 411–421, 2006.pubmed.ncbi.nlm.nih.gov/16788917
  18. [22]R. Bhatia, T. Jain, and Y. Lim, “On the Bures–Wasserstein distance between positive definite matrices,” Expositiones Mathematicae, vol. 37, no. 2, pp. 165–191, 2019.arxiv.org/abs/1712.01504
  19. [23]P. T. Fletcher, S. Venkatasubramanian, and S. Joshi, “The geometric median on Riemannian manifolds, with application to robust atlas estimation,” NeuroImage (Orlando, Fla.), vol. 45, no. 1, pp. S143–S152, 2009.DOI
  20. [24]Y. Thanwerdas and X. Pennec, “The geometry of mixed-Euclidean metrics on symmetric positive definite matrices,” Differential Geometry and its Applications, vol. 81, p. 101867, 2022.arxiv.org/abs/2111.02990
  21. [25]———, “O(n)-invariant Riemannian metrics on SPD matrices,” Linear Algebra and Its Applications, vol. 661, pp. 163–201, 2023.arxiv.org/abs/2109.05768
  22. [26]I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in 7th International Conference on Learning Representations, 2019.
  23. [27]A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala, H. Wallach, H. Larochelle, A. Beygelzimer, F. d’Alché Buc, E. Fox, and R. Garnett, “PyTorch: An imperative style, high-performance deep learning library,” in 33rd Annual Conference on Neural Information Processing Systems, NeurIPS 2019, vol. 32. Neural Information Processing Systems, 2019.arxiv.org/abs/1912.01703
  24. [28]McMaster University Adaptive Systems Lab, “McMaster IPIX radar sea clutter database (Dartmouth 1993 & Grimsby 1998),” http://some.ec.mcmaster.ca/ipix/, accessed 2026-07-19.
  25. [29]F. Gini, M. V. Greco, M. Diani, and L. Verrazzani, “Performance analysis of two adaptive radar detectors against non-Gaussian real sea clutter data,” IEEE Transactions on Aerospace and Electronic Systems, vol. 36, no. 4, pp. 1429–1439, 2000.
  26. [30]K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2016, pp. 770–778.arxiv.org/abs/1512.03385
  27. [31]I. Katsman, E. Chen, S. Holalkere, A. Asch, A. Lou, S. N. Lim, and C. De Sa, “Riemannian residual neural networks,” in Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, pp. 63 502–63 514.arxiv.org/abs/2310.10013
  28. [32]A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, J. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems, vol. 30. Neural Information Processing Systems, 2017.arxiv.org/abs/1706.03762
  29. [33]M. Seraphim, A. Lechervy, F. Yger, L. Brun, and O. Etard, “Structure-preserving transformers for sequences of SPD matrices,” in 2024 32nd European Signal Processing Conference, 2024, pp. 1451–1455.arxiv.org/abs/2309.07579

Paper details

vlnc:W_h5v2xdkr
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents