betaFeedback

WorksPhysics and Astronomy

Quantum-Inspired Robust and Scalable SAR Object Classification

Abstract

SAR image classification naturally has to deal with huge noise and a high dynamic range particularly requiring robust classification models. Additionally, the deployment of these models on edge devices, such as drones and military aircraft, requires a careful balance between model size and classification accuracy. This study explores the potential of tensor networks to meet these robustness requirements, specifically evaluating their resilience to data poisoning. Unlike previous works that concentrated on conventional neural networks for SAR object detection, this research focuses on the robustness and model reduction capabilities of tensor networks in object classification. Our findings indicate that tensor networks are adept at addressing both the challenges of robustness and the need for model efficiency, thereby contributing valuable insights to the ongoing discourse in radar applications and deep learning methodologies in general.

Introduction

The success of deep learning [1, 2, 3, 4, 5] made it also popular for radar applications such as micro-Doppler classification [5, 6, 7, 8, 9] or object detection in synthetic aperture radar (SAR) images [10, 11, 12, 13, 14, 15]. The huge dynamic range and inherent noisy nature of SAR images require a particular focus on the topic of robustness. Furthermore, deployment on edge devices like drones or fighters necessitates a reduction in model size while maintaining a high classification rate. Here, we investigate, how tensor networks (TN) can solve the robustness requirements. We evaluate the robustness to data poisoning using the popular MSTAR SAR classification task [16, 17]. Its moderate size while having a huge target variability, lends itself perfectly to assess the various aspects of model reduction and robustness. We do not focus on the task of SAR image formation (see [18] for a recent overview), but on the aspects of robustness and model reduction of TN to complement our recent investigations [10, 11]. We demonstrate, that TN are well suited to address both challenges.

Tensor Networks

Tensor networks are a mathematical tool to deal with high dimensional data structures that first became popular in the study of quantum many-body systems [19, 20]. Recently, they have been applied to other fields such as optimization [21] and data science, in particular machine learning [22, 23], as an alternative to neural networks. In this section, we briefly introduce the basic notions of TN and how they can be employed for supervised learning, followed by an information theoretical perspective of TN to illustrate some of their advantages over neural networks.

Formal definition

A tensor network is a tuple (T,V)(\mathcal{T},\mathcal{V}) of a set of tensors T={Tn}n=1N\mathcal{T} = \{T_n\}_{n=1}^{N} and a set of links V={vm}m=1M\mathcal{V} = \{v_m\}_{m=1}^{M}. Each tensor Tn∈Rd1×d2×⋯×drnT_n \in\mathbb{R}^{d_1 \times d_2 \times\cdots\times d_{r_n}} has a rank rnr_n and a set of indices In={i1,…,irn}I_n = \{i_1,\ldots,i_{r_n}\} with dimensions d1,…,drnd_1,\ldots,d_{r_n}, respectively1. Each link vmv_m is an index that is shared by exactly two tensors Ti,TjT_i,T_j, such that vm=Ii∩Ijv_m = I_i \cap I_j and vm∉Ikv_m \notin I_k for k≠i,jk \ne i,j. Each link can also be interpreted as the edge of a graph, where the nodes are the tensors and the edges em={Ti,Tj}e_m = \{T_i,T_j\} are given by the tensors that share the index vmv_m. Introducing the convention that indices that appear twice are implicitly summed over, each link also defines a tensor contraction. By summing over all links vmv_m, one can contract the whole tensor network, to obtain a single tensor TI=∑v1…vm(T1)I1…(TN)InT_I = \sum_{v_1 \ldots v_m}(T_1)_{I_1}\ldots(T_N)_{I_n}, that consists of all the indices I=⋃n=1NIn∖VI = \bigcup_{n=1}^{N} I_n \setminus\mathcal{V} that were not contracted and has therefore rank R=∣I∣R = |I|. The tensor network (T,V)(\mathcal{T},\mathcal{V}) encodes the same information as the tensor T\mathcal{T}, but possibly with fewer parameters. The crucial degrees of freedom in a TN are the dimensions χm=dim⁡vm\chi_m = \dim v_m of the links, that control the number of parameters needed to represent the TN. If these so-called bond dimensions are bounded and not exponentially large with respect to the rank rr, the TN is essentially a low-rank decomposition of the original tensor and requires fewer parameters to store the same amount of information. In the context of quantum physics, the tensor TT typically encodes the state of a many-body systems with gapped local Hamiltonians [24], requiring exponentially fewer parameters than the full state vector. This allows a numerical representation which facilitates various computations for large systems which would otherwise be impossible.

Tensor Network geometries

Depending on the internal structure of a tensor’s information content, different graph structures of the tensor network yield better compression rates than others. The most prominent TN architecture is the class of matrix product states (MPS) [19], also referred to as tensor trains (TT), which is based on a linear graph. The MPS decomposition of a tensor TT with rank rr is given by

Ti1⋯ir=(T1)v1i1(T2)v1v2i2(T3)v2v3i3⋯(Tr)vr−1ir(1)T^{i_1\cdots i_r} = (T_1)^{i_1}_{v_1}(T_2)^{i_2}_{v_1v_2}(T_3)^{i_3}_{v_2v_3}\cdots(T_r)^{i_r}_{v_{r-1}} \tag*{(1)}

where indices that appear twice, the links, are implicitly summed over.2 The MPS/TT formalism is powerful, but limited to a certain class of high rank tensors that can be described efficiently, i.e., with bounded bond dimension. In the context of quantum physics, MPS/TT are well suited for 1D systems with local correlations, but fail to describe states of higher dimensional lattices. A generalization of the MPS format is the tree tensor network (TTN), also known as the hierarchical Tucker decomposition (HT), which is based on a binary tree graph. In Figure 1, the graphs of an MPS and a TTN are depicted for a small example. This graphical representation is identical to the Penrose notation of the tensor contractions and provides a simple but powerful way of depicting TN. Both MPS and TTN are loop-free TN, as their graph does not contain closed loops. TN geometries that do contain closed loops [25, 26, 27] can represent a larger class of states efficiently (with bounded bond dimension), but their usage in practice remains challenging, which is why we focus on the TTN in this work as the most flexible loop-free TN.

Example of the graph of an MPS and a TTN encoding a rank-8 tensor

Figure 1. Example of the graph of an MPS (a) and a TTN (b) encoding a rank-8 tensor. The open edges correspond to the indices that are not contracted.

Supervised Learning with Tensor Networks

Supervised learning with tensor networks is based on the decision function

f(x)=WΦ(x),(2)f(x) = W\Phi(x), \tag*{(2)}

for a sample x∈X⊆Rnx \in\mathcal{X} \subseteq\mathbb{R}^{n}, where Φ:X→Rd1×⋯×dn\Phi: \mathcal{X} \to\mathbb{R}^{d_1\times\cdots\times d_n} is the feature map and W∈Rd1×⋯×dn×dlW \in\mathbb{R}^{d_1\times\cdots\times d_n\times d_l} the weight tensor, with dld_l being the number of labels. When the two tensors are contracted, the final result is a vector of dimension dld_l. Similar to kernel methods, non-linear decision boundaries can be achieved through the feature map Φ\Phi that maps the data to an exponentially higher dimensional space. While classical kernel methods rely on the kernel-trick to deal with the high dimensionality, the core idea of tensor network machine learning (TNML) is to encode both Φ(x)\Phi(x) and WW as tensor networks [22]. Starting with the feature map, it is typically defined as a tensor product

Φ(x)=ϕ1(x1)⊗ϕ2(x2)⊗⋯⊗ϕn(xn)(3)\Phi(x) = \phi_1(x_1) \otimes\phi_2(x_2) \otimes\cdots\otimes\phi_n(x_n) \tag*{(3)}

of local feature maps ϕi:R→Hi\phi_i : \mathbb{R} \to\mathcal{H}_i that map a single feature xix_i to a feature vector of a local Hilbert space Hi\mathcal{H}_i. A prominent choice for the local feature map is the spin map ϕ(xi)=(cos⁡(xiπ/2),sin⁡(xiπ/2))T\phi(x_i) = (\cos(x_i\pi/2),\sin(x_i\pi/2))^{\mathsf{T}} with xi∈[0,1]x_i \in[0,1], that maps each feature to a two-dimensional vector that can be interpreted as the state vector of a spin-1/2 particle in a Hilbert space Hi\mathcal{H}_i with dimension two. In this setting, one can not only interpret the input data as the state of a many-body-quantum system, but also the weight tensor WW. This allows us to analyze the weight tensor, i.e., the model itself, with tools from quantum information theory, which will be discussed in more detail in the next section.

In the above definition, the weight tensor WW is exponentially large in the number of features. In order to be able to represent it numerically and optimize it, we encode it directly as a tensor network (W,V)(\mathcal{W},\mathcal{V}) with tuneable bond dimensions, without ever having to construct the full tensor. This introduces bias, but enables us to use variations of efficient optimization algorithms for quantum systems like the density matrix renormalization group (DMRG) algorithm [19] or Riemannian optimization [28]. Furthermore, the bond dimensions of the TN are directly related to the expressivity of the model, and can be tuned to avoid over-fitting. This framework has been successfully applied to standard computer vision datasets such as the MNIST or fashion-MNIST, achieving state-of-the-art results [22, 29, 30].

Information theoretical perspective

The interpretation of a TN as a quantum mechanical state opens the possibility to study it from the perspective of quantum information theory, which studies the information content and correlations of quantum mechanical systems. Applying this framework to a trained TNML model can give us deep insights into the way the model processes information, and based on what criteria decisions are made.

Schmidt decomposition

An important concept of quantum information theory is the Schmidt decomposition. If H=HA⊗HB\mathcal{H} = \mathcal{H}_A \otimes\mathcal{H}_B is a bipartite Hilbert space with dim⁡HA=n\dim\mathcal{H}_A = n and dim⁡HB=m\dim\mathcal{H}_B = m, then the Schmidt decomposition of a vector w∈Hw \in\mathcal{H} is defined as

w=∑i=1rαiui⊗vi(4)w = \sum_{i=1}^{r} \alpha_i u_i \otimes v_i \tag*{(4)}

with orthonormal sets {u1,…,ur}∈HA\{u_1, \ldots, u_r\} \in\mathcal{H}_A and {v1,…,vr}∈HB\{v_1, \ldots, v_r\} \in\mathcal{H}_B, and real, non-negative Schmidt coefficients αi\alpha_i. The Schmidt rank r≤min⁡{n,m}r \leq\min\{n,m\} is the number of non-zero Schmidt coefficients. The Schmidt coefficients are unique up to reordering, and in the following we assume that they are ordered decreasingly. Note that this is essentially a reformulation of the singular value decomposition (SVD) in another context. For TN without loops, like MPS or TTN, the Schmidt decomposition can be computed efficiently through a series of local SVD and QR decompositions on the tensors T\mathcal{T} (see e.g., [19 Sec. 4.1.3] for details).

Entanglement entropy

Analogous to the Shannon entropy in classical information theory, the bipartite entanglement entropy, or von Neumann entropy, is the basic measure of information in a quantum mechanical system. It quantifies the entanglement over a bipartition A;BA; B of a quantum mechanical system H\mathcal{H} into two complimentary subsystems HA\mathcal{H}_A and HB\mathcal{H}_B with H=HA⊗HB\mathcal{H} = \mathcal{H}_A \otimes\mathcal{H}_B. It can be expressed in terms of the Schmidt coefficients αi\alpha_i of the bipartition as

H(A;B)=−∑iαi2log⁡αi2.(5)H(A; B) = -\sum_i \alpha_i^2 \log\alpha_i^2. \tag*{(5)}

Note that this is the Shannon entropy of the squared Schmidt coefficients, which define a probability distribution if the state is normalized. The entanglement entropy quantifies the amount of information that is shared between the two subsystems. In the context of machine learning, the weight tensor WW is defined on a product of Hilbert spaces W∈H=H1⊗⋯⊗Hn⊗HlW \in\mathcal{H} = \mathcal{H}_1 \otimes\cdots\otimes\mathcal{H}_n \otimes\mathcal{H}_l, one for each feature and one for the label. This is due to the fact that the feature map Φ\Phi maps each feature xix_i to a local Hilbert space Hi\mathcal{H}_i, and the additional index that corresponds to the output ff can also be interpreted to be a vector of a Hilbert space Hl\mathcal{H}_l. This means that we can compute the entanglement entropy of any bipartition of features and the label efficiently through the Schmidt coefficients of the TN. In particular, the entropy of the bipartition of a single feature with the remaining system is of interest. This quantity tells us how relevant this feature is for the classification, as the entanglement entropy quantifies the amount of information shared between the feature and the rest of the model. Note that the quantity is computed from the model itself, without having to consider any data samples. It is therefore a measure of feature importance that is data independent and exact, i.e., without sampling error.

Compression

Another powerful property of TN is the ability to compress them in a controlled fashion, discarding the least relevant information first. To see how this is done, consider again the Schmidt decomposition (Eq. (4)). Now note that in a loop free TN, each link vmv_m corresponds to a bipartition of the system, with a corresponding Schmidt decomposition. If the bond dimension χ\chi, i.e., the dimension of a link, is larger than the corresponding Schmidt rank rr, then it can be truncated without loosing any information, since the extra dimensions have Schmidt values that are equal to zero. When all bond dimensions are equal to the corresponding Schmidt ranks, the network is said to be of full rank. In the same spirit, the TN can be compressed further by discarding the smallest kk Schmidt values of a bipartition, s.t.

∑i=0k−1αr−i2≤ε,(6)\sum_{i=0}^{k-1} \alpha_{r-i}^{2} \leq\varepsilon, \tag*{(6)}

where ε>0\varepsilon> 0 is a tunable error threshold, and kk is chosen to be as large as possible while still fulfilling this condition. By repeating this procedure for all links vmv_m, always discarding the Schmidt values of the corresponding bipartition such that Eq. (6) is satisfied, the full network is compressed. This expression introduces an error in the representation of the full tensor, but since the Schmidt values correspond to singular values, we know that the error of this compression, with respect to the Frobenius norm, is given by the sum of the squared Schmidt-/ singular values that were discarded.

Numerical experiments

For numerical evaluations, we used the well established MSTAR dataset [16, 17]. The MSTAR Moving and Stationary Target Acquisition and Recognition dataset is a publicly available benchmark developed by the U.S. DARPA and AFRL for automatic target recognition using Synthetic Aperture Radar imagery. It contains high-resolution SAR images of military ground vehicles captured under varying depression angles and configurations. The dataset includes multiple target classes, such as tanks and armored personnel carriers, with labeled imagery for training and testing recognition algorithms. It has been used in a wide range of applications (e.g., [31, 32, 33, 34]). Here, we employ the main 8 target classes. The images were compressed to 32×3232 \times32 pixels, and the dataset was split randomly into a training set containing 70% (6626) of the samples and a test set with the remaining 30% (2840).

Robustness against data poisoning

In the MSTAR dataset, all targets are located at the center of the images and are surrounded by a background of clutter. A common pitfall of this dataset is that machine learning models learn to classify using information in the background, rather than the targets themselves. This will lead to poor performance on real data, where the background might be different. With TN models, we can easily assert that the model learned to use information of the targets by looking at the feature entropies of the model. In Figure 2, one can see the feature entropies of a TTN model that was trained on the MSTAR dataset, reaching a test accuracy of 99.05%. In the area where the target is located, the feature entropies are significantly higher than in the surrounding area. Since the feature entropies are a measure of a feature’s importance for the classification, this tells us that the model learned to use the information of the targets, while mostly neglecting the background. To demonstrate the effectiveness of this method, we now artificially correlate a single pixel (feature) located in the background with the class for all samples. If kk is the index of the pixel we modify, then the feature xi,kx_{i,k} of sample ii is replaced by

Figure 2. Feature entropies of the TTN model trained on the original MSTAR dataset.

xˉi,k=8−yi10⋅X(7)\bar{x}_{i,k} = \frac{8-y_i}{10} \cdot X \tag*{(7)}

where XX represents multiplicative, Gaussian distributed noise X∼N(1,10−4)X \sim\mathcal{N}(1, 10^{-4}) and yi={0,1,…,7}y_i = \{0, 1, \ldots, 7\} is the numerical value assigned to the class of sample ii. The modification was chosen such that the modified pixel blends well with the clutter background, making it difficult to detect the change by simply looking at the data. The model trained on this data reaches 99.3% test accuracy on the modified test data, but only 92.68% on the original (clean) test data. This suggests that the model is relying on the information encoded in the modified pixel, but without access to the clean data this cannot be tested. However, by looking at the feature entropies again (Figure 3), we can immediately see that something is wrong, as the feature entropy of the modified pixel is by far the largest. Instead of replacing a single feature, we can also correlate the whole background to the target class by injecting multiplicative Gaussian noise (speckle) that depends on the target class. The noise is only applied to the clutter background D\mathcal{D}, to not affect the targets and make the modification less obvious (see Figure 4). The modified features xˉi,k\bar{x}_{i,k} are given by

Figure 3. Feature entropies of the TTN model trained on a poisoned dataset where one single pixel in the background was correlated to the target class.

Figure 4. Illustration of the noise that was applied to the background. The difference between the original image (a) and the poisoned image (b) is barely visible. c) The speckle noise that was injected to the background region D\mathcal{D}. The region in the center, where the target is located, is left unchanged. This specific noise pattern was applied to all samples of this class, and a different pattern to the other classes, respectively, but all generated from the same normal distribution.

xˉi,k={xi,k⋅nyi,k,k∈Dxi,k,else(8)\bar{x}_{i,k} = \begin{cases} x_{i,k} \cdot n_{y_i,k}, & k \in\mathcal{D} \\ x_{i,k}, & \text{else} \end{cases} \tag*{(8)}

with ny,k∼N(1,0.02)n_{y,k} \sim\mathcal{N}(1, 0.02). The TTN classifier trained on this data achieved a test accuracy of 99.86% on the poisoned data, and 82.18% on the clean test data. In Figure 5, one can see the feature entropies of this model. Clearly, the model is relying heavily on the modified background, which explains the significant drop in accuracy on the clean test data, where the patterns in the background that the model learned are not present.

Figure 5. Feature entropies of the model TTN trained on a poisoned dataset where the background was strongly correlated to the target class.

Model compression

To demonstrate the power of tensor network compression in the context of machine learning, we compress the model trained on the MSTAR dataset with gradually increasing error threshold ε\varepsilon, which defines the maximal sum of squared singular values that are discarded (Eq. (6)). In Figure 6, one can see how the number of parameters of the model, which is proportional to the inference time, and the test accuracy change as the compression tolerance ε\varepsilon is increased and the model gets compressed more strongly. This particular model can be compressed up to a threshold of ε=10−3\varepsilon= 10^{-3} without a loss of accuracy, while the number of parameters is decreasing. In this loss-free domain, the model can be compressed up to a compression ratio

Figure 6. Comparison of the number of parameters in the model and the test accuracy of the compressed model with increasing compression threshold ε\varepsilon.

r=compressed sizeuncompressed sizer = \frac{\text{compressed size}}{\text{uncompressed size}}

of about 75%. With higher tolerances, accuracy gradually decreases along with the number of parameters. For r≈47%r \approx47\%, the test accuracy is still at 97.11%. Since inference time is proportional to the number of parameters, there is a trade-off between accuracy and inference time/ model size in this domain. Similar behavior can be observed for all TNML learning models, although the concrete relationship between the compression tolerance, the accuracy and the number of parameters depends on many factors such as the quality of the data or the model’s ability to generalize.

Summary

This study investigated the application of tensor networks in radar technologies, particularly for SAR object classification. The inherent challenges posed by the vast dynamic range and noise in SAR imagery necessitated a strong emphasis on robustness. We evaluated the resilience of tensor networks to data poisoning, including the presence of poisonous pixels, through the MSTAR SAR classification task. This task was particularly suitable due to its moderate dataset size and substantial target variability, allowing for a comprehensive assessment of model reduction and robustness. The findings revealed that tensor networks achieved significant model reduction while maintaining high classification accuracy, making them ideal for deployment in resource-constrained environments. Furthermore, this research emphasized the robustness and model reduction capabilities of tensor networks in object classification, indicating that they were well-suited to tackle the dual challenges of enhancing robustness and achieving model efficiency. Overall, the results contributed valuable insights to the ongoing discourse in radar applications and deep learning methodologies, since model reduction and robustness are becoming more and more important as models are deployed to an ever wider set of applications.

Outlook

The results presented here have important practical consequences for the enhancement of object detection pipelines. First, the ability to tune models by balancing computing power, desired inference time, and model performance facilitates context-dependent object identification. That is, models can be tweaked depending on the importance of high accuracy and whether or not longer computation times are acceptable. Second, the ability to identify data poisoning lends itself to a degree of explainability, which is also important in practice. Both of these points can be seen, for example, in a military setting, where being able to adapt a classification model to the context could result in operational advantages. On top of this, having the ability to identify adversarial attacks or when a mission-critical model is making ill-informed decisions could assist a human operator in making important decisions quickly and confidently. While the data used here demonstrate the potential of the proposed approach in the defense domain, the approach proposed here is certainly not restricted to this setting. An example target classification pipeline could look as follows: Spaceborne radar systems collect data over specific areas of interest. A TNML model would then be selected based on the operational context, and the resulting classifications, combined with feature importance information, this would give additional information to a human actor to aid their decision-making. In general, this work follows the trend of incorporating quantum and quantum-inspired algorithms in radar, and further work could include testing data sets from different domains and comparing results to traditional machine learning approaches.

Acknowledgements

This work is part of the QUA-SAR project by the DLR Quantum Computing Initiative (DLR QCI) and is funded by the Federal Ministry of Research, Technology and Space (BMFTR). https://qci.dlr.de/qua-sar

References

  1. [1]D. C. Ciresan et al. “Deep Big Simple Neural Nets for Handwritten Digit Recognition”. In: Neural Computation 22.12 (2010), pp. 3207–20.DOI
  2. [2]A. Graves and J. Schmidhuber. “Offline Handwriting Recognition withMultidimensional Recurrent Neural Networks”. In: Advances in Neural Information Processing Systems 21. NIPS. Curran Associates, Inc., 2009, pp. 545–552.
  3. [3]A. Krizhevsky, I. Sutskever, and G. E. Hinton. “ImageNet Classification with Deep Convolutional Neural Networks”. In: NIPS. 2012.
  4. [4]Y. LeCun et al. “Backpropagation Applied to Handwritten Zip Code Recognition”. In: Neural Computation 1.4 (1989), pp. 541–51.DOI
  5. [5]C. Neumann and T. Brosch. “Deep Learning Approach for Radar Applications”. In: 21st International Radar Symposium (IRS). 2020, pp. 28–30.DOI
  6. [6]F. Hoffmann et al. “Micro-Doppler based detection and tracking of UAVs with multistatic radar”. In: 2016 IEEE Radar Conference (RadarConf). 2016, pp. 1–6.
  7. [7]A. Hanif et al. “Micro-Doppler Based Target Recognition With Radars: A Review”. In: IEEE Sensors Journal 22.4 (2022), pp. 2948–2961.DOI
  8. [8]R. Petervari et al. “A Realtime Micro-Doppler Detection, Tracking and Classification System for the 94 GHz FMCW Radar System DUSIM”. In: 2019 16th European Radar Conference (EuRAD). 2019, pp. 193–196.
  9. [9]R. Amiri and A. Shahzadi. “Micro-Doppler based target classification in ground surveillance radar systems”. In: Digital Signal Processing 101 (2020), p. 102702.DOI
  10. [10]T. Brosch and C. Neumann. “Automatic Target Recognition on High Resolution SAR”. In: 21st International Radar Symposium (IRS). IEEE, 2021, pp. 1–6.
  11. [11]M. Schwaiger et al. “Ultrafast Object Detection on High Resolution SAR”. In: International Radar Symposium. 2022, pp. 1–5.DOI
  12. [12]Y. Almalioglu et al. SAR Object Detection with Self-Supervised Pretraining and Curriculum-Aware Sampling. 2025.arxiv.org/abs/2504.13310
  13. [13]Z. Wang et al. “Object Detection on SAR Images Via YOLOv10 and Integrated ACmix Attention Mechanism”. In: 2024 6th International Conference on Robotics, Intelligent Control and Artificial Intelligence (RICAI). 2024, pp. 756–760.DOI
  14. [14]M. Warner et al. “Maritime Object Detection in Synthetic Aperture Radar (SAR) Imagery Using Deep Learning”. In: AGU Fall Meeting Abstracts. Vol. 2020. AGU Fall Meeting Abstracts. Dec. 2020, G004-0022, G004–0022.
  15. [15]A. Pulella et al. “Generalization in deep learning-based aircraft classification for SAR imagery”. In: ISPRS Journal of Photogrammetry and Remote Sensing 218 (2024), pp. 312–323.DOI
  16. [16]E. R. Keydel, S. W. Lee, and J. T. Moore. “MSTAR Extended Operating Conditions: A Tutorial”. In: vol. 2757. International Society for Optics and Photonics. SPIE, 1996, pp. 228–42.DOI
  17. [17]T. Ross et al. “Standard SAR ATR Evaluation Experiments Using the MSTAR Public Release Data Set”. In: Algorithms for Synthetic Aperture Radar Imagery V. Vol. 3370. SPIE, 1998, pp. 566–73.DOI
  18. [18]H. Cruz et al. “A Review of Synthetic-Aperture Radar Image Formation Algorithms and Implementations A Computational Perspective”. In: Remote Sensing 14.1258 (2022), pp. 1–30.DOI
  19. [19]U. Schollwöck. “The density-matrix renormalization group in the age of matrix product states”. In: Annals of Physics 326.1 (Jan. 2011), 96–192.DOI
  20. [20]S. Montangero. Introduction to Tensor Network Methods. Springer International Publishing, 2018.
  21. [21]S. Cavinato et al. “Optimizing radiotherapy plans for cancer treatment with Tensor Networks”. In: Physics in Medicine & Biology 66.12 (June 2021), p. 125015.DOI
  22. [22]E. M. Stoudenmire and D. J. Schwab. Supervised Learning with Quantum-Inspired Tensor Networks. 2017.arxiv.org/abs/1605.05775
  23. [23]T. Felser et al. “Quantum-inspired machine learning on high-energy physics data”. In: npj Quantum Information. 1st ser. 7 (2021).
  24. [24]M. B. Hastings. “Solving gapped Hamiltonians locally”. In: Physical Review B 73.8 (Feb. 2006).DOI
  25. [25]G. Evenbly and G. Vidal. “Algorithms for entanglement renormalization”. In: Physical Review B 79.14 (Apr. 2009).
  26. [26]T. Felser, S. Notarnicola, and S. Montangero. “Efficient Tensor Network Ansatz for High-Dimensional Quantum Many-Body Problems”. In: Physical Review Letters 126.17 (Apr. 2021).DOI
  27. [27]F. Verstraete et al. “Criticality, the Area Law, and the Computational Power of Projected Entangled Pair States”. In: Physical Review Letters 96.22 (June 2006).pubmed.ncbi.nlm.nih.gov/16803296
  28. [28]M. Willner, M. Trenti, and D. Liebiedz. Riemannian Optimization on Tree Tensor Networks with Application in Machine Learning. 2025.arxiv.org/abs/2507.21726
  29. [29]I. Glasser, N. Pancotti, and J. I. Cirac. From probabilistic graphical models to generalized tensor networks for supervised learning. 2019.arxiv.org/abs/1806.05964
  30. [30]E. M. Stoudenmire. “Learning relevant features of data with multi-scale tensor networks”. In: Quantum Science and Technology 3.3 (Apr. 2018), p. 034003.DOI
  31. [31]E. Blasch et al. “Review of Recent Advances in AI/ML Using the MSTAR Data”. In: Algorithms for Synthetic Aperture Radar Imagery XXVII. Vol. 11393. SPIE, 2020, p. 11393OC.DOI
  32. [32]J. Fein-Ashley et al. Benchmarking Deep Learning Classifiers for SAR Automatic Target Recognition. 2023.
  33. [33]R. Schumacher and K. Rosenbach. “ATR of Battlefield Targets by SAR – Classification Results Using the Public MSTAR Dataset Compared with a Dataset by QinetiQ, UK”. In: Research and Technology Organization, Sensors and Electronics Technology Panel (Symposium). RTO–MP–SET–080, 2004, pp. 31–1–31–12.
  34. [34]A. Singh and V. K. Singh. Exploring Deep Learning Methods for Classification of SAR Images: Towards NextGen Convolutions via Transformers. 2023.

Paper details

vlnc:W_z2ugrk7w
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.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents