Abstract:Synthetic aperture radar technology is crucial for high-resolution imaging under various conditions; however, the acquisition of real-world synthetic aperture radar data for deep learning-based automatic target recognition remains challenging due to high costs and data availability issues. To overcome these challenges, synthetic data generated through simulations have been employed, although discrepancies between synthetic and real data can degrade model performance. In this study, we introduce a novel framework, soft segmented randomization, designed to reduce domain discrepancy and improve the generalize ability of synthetic aperture radar automatic target recognition models. The soft segmented randomization framework applies a Gaussian mixture model to segment target and clutter regions softly, introducing randomized variations that align the synthetic data's statistical properties more closely with those of real-world data. Experimental results demonstrate that the proposed soft segmented randomization framework significantly enhances model performance on measured synthetic aperture radar data, making it a promising approach for robust automatic target recognition in scenarios with limited or no access to measured data.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is the performance degradation of deep - learning models in Synthetic Aperture Radar Automatic Target Recognition (SAR - ATR) due to the domain differences between synthetic data and measured data. Specifically: 1. **Difficulty in data acquisition**: Acquiring Synthetic Aperture Radar (SAR) data in the real world is very complex and costly. Especially in military applications, data of specific military targets such as foreign tanks or airplanes can hardly be directly obtained. 2. **Differences between synthetic data and measured data**: Although synthetic data can be generated by simulation to train deep - learning models, there are significant differences between synthetic data and measured data. These differences are mainly reflected in the differences in background clutter and target features, resulting in poor performance of the model on measured data. To solve these problems, the paper proposes a new framework - Soft Segmented Randomization (SSR), aiming to reduce domain differences and improve the generalization ability of the model in the following ways: - **Soft segmentation**: Use the Gaussian Mixture Model (GMM) to perform soft segmentation on synthetic SAR images, separating the target area and the clutter area. - **Randomization**: Introduce random changes in the segmented areas, adjust the statistical properties (such as mean and variance) of the clutter area to make them closer to the statistical properties of the measured data. - **Preserving key features**: During the adjustment process, try to preserve the key domain - independent features in the target area, such as high - intensity scattering points and large - shape features. Through these methods, the SSR framework can significantly improve the performance of deep - learning models trained on synthetic data on measured data, thereby reducing the dependence on measured data and improving the robustness and generalization ability of the model. ### Main contributions 1. **Identifying domain differences at the image level**: Analyze in detail the main differences between synthetic data and measured data in SAR - ATR, especially the changes in clutter mean and variance and target features. 2. **Developing the SSR framework**: Propose the SSR framework using GMM for soft segmentation, reducing clutter reflectivity and speckle noise differences through targeted randomization while preserving the key features in the target area. 3. **Enhancing domain generalization in SAR - ATR**: The SSR framework significantly improves the generalization ability of deep - learning models trained only on synthetic data on measured data. 4. **Practical applications**: Provide a practical solution suitable for SAR - ATR systems where it is difficult or impossible to obtain measured data, effectively bridging the gap between synthetic data and measured data. Through these improvements, the SSR framework provides new ideas and methods for the performance improvement and reliability of SAR - ATR systems.

Soft Segmented Randomization: Enhancing Domain Generalization in SAR-ATR for Synthetic-to-Measured

IRASNet: Improved Feature-Level Clutter Reduction for Domain Generalized SAR-ATR

Single-Scene SAR Image Data Augmentation Based on SBR and GAN for Target Recognition

Target Recognition in SAR Images by Deep Learning with Training Data Augmentation

Semi-Supervised SAR ATR Framework with Transductive Auxiliary Segmentation

Fast sparse adversarial attack for synthetic aperture radar target recognition

Gradual Domain Adaptation with Pseudo-Label Denoising for SAR Target Recognition When Using Only Synthetic Data for Training

SAR Target Recognition Based on Probabilistic Meta-Learning

Adaptive Convolutional Subspace Reasoning Network for Few-Shot SAR Target Recognition

Progressive Multi-Level Alignments for Semi-Supervised Domain Adaptation SAR Target Recognition Using Simulated Data

ATGAN: A SAR Target Image Generation Method for Automatic Target Recognition

Weighted residual network for SAR automatic target recognition with data augmentation

Exploring Reinforced Class Separability and Discriminative Representations for SAR Target Open Set Recognition

Improving SAR Automatic Target Recognition via Trusted Knowledge Distillation From Simulated Data

SAR ATR for Limited Training Data Using DS-AE Network

Transductive Prototypical Attention Reasoning Network for Few-Shot SAR Target Recognition

An Entropy-Awareness Meta-Learning Method for SAR Open-Set ATR

Open Set Recognition and Category Discovery Framework for SAR Target Classification Based on K-Contrast Loss and Deep Clustering

Robust SAR Automatic Target Recognition Via Adversarial Learning

SAR ATR Method with Limited Training Data via an Embedded Feature Augmenter and Dynamic Hierarchical-Feature Refiner

Multilevel Adaptive Knowledge Distillation Network for Incremental SAR Target Recognition