Flow-SAR: Cross-Domain SAR Image Translation Using Flow Models

In synthetic aperture radar (SAR) automatic target recognition (ATR), the discrepancy between the source domain and the target domain affects the model’s generalization performance. Existing generative adversarial network (GAN)-based image translation methods have effectively alleviated this problem, but GAN itself inherently suffers from unstable training and mode collapse. To overcome these drawbacks, we propose Flow-SAR, a two-stage SAR image cross-domain translation framework based on invertible flow models. Flow-SAR maps source domain and target domain images into a shared latent space with a learnable Gaussian prior and explicitly aligns feature distributions via latent vector translation. In the first stage, a global translation vector is estimated by aligning the mean latent representations of both domains to achieve coarse distribution alignment. In the second stage, a class-specific refinement mechanism iteratively generates high-confidence pseudo-labels for target samples to estimate class-wise translation vectors, thus enhancing intra-class consistency and improving target recognition accuracy. Experiments on the SAMPLE dataset demonstrate that Flow-SAR improves the visual realism of transformed SAR images and effectively enhances target recognition performance.