SHFGAN: SARImage Harmonization Fusion with Generative Adversarial Network

Although significant progress has been achieved in synthesizing and simulating SAR (Synthetic Aperture Radar) images using generative models, the random generation of target characteristics and target disappearance during the synthesis process remain major challenges due to the unique nature of SAR images. To address these issues while ensuring harmonious fusion of target patches and backgrounds, this paper proposes a generative model named SHFGAN (SAR Image Harmonization Fusion Generative Adversarial Network). Built upon a multi constraint generative adversarial network (GAN) framework, the generator of SHFGAN employs a ViT (Vision Transformer) model. By introducing attention mechanisms to enhance feature interaction and integrating a semantic embedding module to enforce external semantic constraints, the model effectively resolves unnatural target-background fusion and random generation of target characteristics. Additionally, a multi-task discriminator design, including a realism discriminator and a mask segmentation discriminator, is adopted to respectively constrain the authenticity of generated images and their semantic segmentation results. Through adversarial training and multi constraint loss functions, the problem of target disappearance is mitigated. Quantitative and qualitative experiments on the SARDet-100K, SSDD and AIR-SARship datasets demonstrate that SHFGAN achieves remarkable performance in SAR target patch-background fusion tasks, outperforming common generative models and reaching state-of-the-art levels.