Learning Capsules for SAR Target Recognition

Yunrui Guo, Zongxu Pan, Meiming Wang, Ji Wang, Wenjing YangIEEE JSTARS 2020 · OpenReview

Abstract

Deep learning has been successfully utilized in synthetic aperture radar (SAR) automatic target recognition tasks and obtained state-of-the-art results. However, current deep learning algorithms do not perform well when SAR images are occluded, noisy, or with a great depression angle variance. This article proposes a novel method, SAR capsule network, to achieve the accurate and robust classification of SAR images without significantly increasing network complexity. Specifically, we develop a convolutional neural network extension based on Hinton's capsule network to capture spatial relationships specialized in classification between different entities in a SAR image. The SAR capsules are learned by a vector-based full connection operation instead of the traditional routing process, which not only alleviates the computational burden but also improves recognition accuracy. For occlusion, additive noise, and multiplicative noise tests, SAR capsule network shows superior robustness compared with typical convolution neural networks. When missing training data in a certain aspect angle range or existing a large depression angle variance between training data and test data, the proposed network achieves better performance than the existing works and reveals some competitive advantages in several test scenarios.

Introduction

Synthetic Aperture Radar (SAR) is a critical technology for all-weather, 24/7 Earth observation. However, automatic target recognition (ATR) in SAR imagery remains a significant challenge due to the high sensitivity of radar signatures to changes in depression and aspect angles. While standard Convolutional Neural Networks (CNNs) have achieved success in many computer vision tasks, they often struggle with SAR data. This is primarily due to the loss of spatial relationship information caused by pooling layers and a lack of robustness under Extended Operating Conditions (EOCs), such as occlusion or significant noise.

Methodology: The SAR Capsule Network

To address these limitations, we propose the SAR Capsule Network. Unlike traditional CNNs that use scalar-valued neurons, this architecture utilizes "capsules"—groups of neurons whose output is a vector. The length of the vector represents the probability of a target's presence, while its orientation encodes instantiation parameters such as position, size, and pose.

The architecture of the proposed SAR Capsule Network, featuring an encoder-decoder structure and vector-based capsules.

A key innovation in our approach is the Vector-based Full Connection (VFC) mechanism. Traditional capsule networks rely on an iterative "dynamic routing" process, which is computationally intensive. VFC replaces this with learnable connection weights, improving both training speed and convergence stability. Additionally, the network incorporates a reconstruction regularizer—a decoder that attempts to recreate the input image from the capsule vectors. This forces the model to learn the physical essence of the target rather than superficial pixel patterns.

Experimental Results and Robustness

The model was evaluated using the MSTAR dataset, covering both Standard Operating Conditions (SOC) and Extended Operating Conditions (EOC). Under SOC, the SAR Capsule Network achieved a state-of-the-art accuracy of 99.79%, outperforming established models like ResNet and A-ConvNet.

The network's robustness was particularly evident in occlusion and noise tests. Even when 36% of target pixels were missing, the model maintained over 70% accuracy. The internal decoder's ability to reconstruct the target's outline from partial data demonstrates that the capsules capture high-level abstract representations of the vehicles.

Visualizing robustness: The network's ability to reconstruct target outlines (c) even when significant portions of the original SAR image (a) are occluded (b).

Furthermore, in EOC-1 tests involving large depression angle variances, the model achieved 98.78% accuracy, significantly surpassing previous multi-view and transfer-learning methods. The model also showed superior resistance to both additive and multiplicative (speckle) noise, which is common in radar imaging.

Conclusion

The SAR Capsule Network provides a robust and efficient solution for SAR target recognition. By preserving spatial relationships through vector-based capsules and utilizing a learnable VFC mechanism, the model achieves high accuracy and remarkable resilience to the challenging conditions inherent in radar imaging. Future work will explore more complex cluttered environments and the integration of multi-agent teamwork concepts to refine capsule interactions.

Summary updated

Abstract

Deep learning has been successfully utilized in synthetic aperture radar (SAR) automatic target recognition tasks and obtained state-of-the-art results. However, current deep learning algorithms do not perform well when SAR images are occluded, noisy, or with a great depression angle variance. This article proposes a novel method, SAR capsule network, to achieve the accurate and robust classification of SAR images without significantly increasing network complexity. Specifically, we develop a convolutional neural network extension based on Hinton's capsule network to capture spatial relationships specialized in classification between different entities in a SAR image. The SAR capsules are learned by a vector-based full connection operation instead of the traditional routing process, which not only alleviates the computational burden but also improves recognition accuracy. For occlusion, additive noise, and multiplicative noise tests, SAR capsule network shows superior robustness compared with typical convolution neural networks. When missing training data in a certain aspect angle range or existing a large depression angle variance between training data and test data, the proposed network achieves better performance than the existing works and reveals some competitive advantages in several test scenarios.
BibTeX
@article{guo2020learning,
  author = {Yunrui Guo and Zongxu Pan and Meiming Wang and Ji Wang and Wenjing Yang},
  title = {{Learning Capsules for SAR Target Recognition}},
  journal = {IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens.},
  year = {2020},
  url = {https://openreview.net/forum?id=2W1TK8Bywu},
}
Figure