Abstract:Sufficient synthetic aperture radar (SAR) target images are very important for the development of researches. However, available SAR target images are often limited in practice, which hinders the progress of SAR application. In this paper, we propose an azimuth-controllable generative adversarial network to generate precise SAR target images with an intermediate azimuth between two given SAR images' azimuths. This network mainly contains three parts: generator, discriminator, and predictor. Through the proposed specific network structure, the generator can extract and fuse the optimal target features from two input SAR target images to generate SAR target image. Then a similarity discriminator and an azimuth predictor are designed. The similarity discriminator can differentiate the generated SAR target images from the real SAR images to ensure the accuracy of the generated, while the azimuth predictor measures the difference of azimuth between the generated and the desired to ensure the azimuth controllability of the generated. Therefore, the proposed network can generate precise SAR images, and their azimuths can be controlled well by the inputs of the deep network, which can generate the target images in different azimuths to solve the small sample problem to some degree and benefit the researches of SAR images. Extensive experimental results show the superiority of the proposed method in azimuth controllability and accuracy of SAR target image generation.

What problem does this paper attempt to address?

The paper attempts to address the issue of insufficient Synthetic Aperture Radar (SAR) target image data. Specifically, the number of available SAR target images in practical applications is limited, which restricts the development and application of SAR technology, especially in fields such as Automatic Target Recognition (ATR). To overcome this problem, the authors propose an Azimuth-Controllable Generative Adversarial Network (GAN) to generate accurate SAR target images with specific azimuth angles. ### Main Issues 1. **Insufficient SAR Image Data**: The existing amount of SAR image data is limited, difficult to obtain, and resource-intensive. 2. **Azimuth Control**: Most existing SAR image generation methods can only generate images from random noise and cannot control the azimuth angle of the generated images. This results in uneven distribution of generated images at certain azimuth angles, affecting the accuracy of research such as target recognition. ### Solution The authors propose an Azimuth-Controllable Generative Adversarial Network (Azimuth-Controllable GAN), which includes three main components: 1. **Generator**: Extracts SAR image features at different azimuth angles through two parallel input blocks and generates SAR images at intermediate azimuth angles through an information fusion block. 2. **Discriminator**: Includes a similarity discriminator and an azimuth predictor. The similarity discriminator distinguishes between generated SAR images and real SAR images to ensure the accuracy of the generated images; the azimuth predictor measures the difference between the azimuth angle of the generated image and the desired azimuth angle to ensure the controllability of the azimuth angle of the generated image. 3. **Loss Function**: By optimizing the loss functions of the generator, discriminator, and predictor, the generated SAR images are made to be close to real images in terms of both similarity and azimuth angle. ### Main Contributions 1. **Proposed an Azimuth-Controllable GAN Framework**: Through adversarial training between the generator and discriminator, and the introduction of the azimuth predictor, the framework learns to extract features from two input SAR images and generate accurate SAR images with intermediate azimuth angles. 2. **Special Topological Structure Design**: The generator adopts dual-input parallel blocks and residual blocks, while the discriminator and predictor use strided convolution and deformable convolution, improving the model's stability and the quality of the generated images. 3. **Experimental Validation**: Experiments on the MSTAR dataset validate the effectiveness of the proposed method in generating high-quality SAR images and improving target recognition performance. ### Experimental Results 1. **Visual Similarity of Generated Images**: The generated SAR images are visually very similar to real images, and the azimuth angles are within the desired range. 2. **Quality Evaluation of Generated Images**: The quality of the generated images is quantitatively evaluated using image similarity metrics. 3. **Improvement in Target Recognition Performance**: The generated SAR images significantly improve target recognition performance under Standard Operating Conditions (SOC) and Extended Operating Conditions (EOC). In summary, this paper effectively addresses the issue of insufficient SAR target image data by proposing an Azimuth-Controllable Generative Adversarial Network, providing a new solution for the generation and application of SAR images.

SAR Target Image Generation Method Using Azimuth-Controllable Generative Adversarial Network

Attribute-Controlled SAR Target Image Generation Driven by Diffusion Models

SAR Image Generation Using Structural Bayesian Deep Generative Adversarial Network

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

Circular synthetic aperture radar sub‐aperture angle information complementation based on azimuth‐controllable generative adversarial network

ISAR Images Generation Via Generative Adversarial Networks

SAR-PeGA: A Generation Method of Adversarial Examples for SAR Image Target Recognition Network

Learning to Generate SAR Images With Adversarial Autoencoder

A-CGAN based transformation from ISAR to optical image

SAR Image Generation Method Using DH-GAN for Automatic Target Recognition

A SAR-to-Optical Image Translation Method Based on Conditional Generation Adversarial Network (cGAN)

A Two-Stage SAR Image Generation Algorithm Based on GAN with Reinforced Constraint Filtering and Compensation Techniques

SAR Image Generation by Integrating Differentiable SAR Renderer with Neural Networks

Automatic Target Recognition for Synthetic Aperture Radar Images Based on Super-Resolution Generative Adversarial Network and Deep Convolutional Neural Network

Enhanced small target detection and color interpretation of synthetic aperture radar images

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

Super-resolution SAR Image Reconstruction Via Generative Adversarial Network

A GAN-based fast focusing method for circular SAR images

Zero-Shot Learning of SAR Target Feature Space with Deep Generative Neural Networks.

Parallel Connected Generative Adversarial Network with Quadratic Operation for SAR Image Generation and Application for Classification

Zero-Shot SAR Target Recognition Based on a Conditional Generative Network with Category Features from Simulated Images