MIRB-Net: Masked Inpainting and Residual Blending for Conditional SAR Target Generation on Backgrounds

Enhancing the environmental generalization of Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) requires diverse target–clutter combinations. Despite abundant target-free clutter, exhaustively capturing multi-background target samples is prohibitive. However, existing generative methods either distort authentic clutter statistics through full-scene reconstruction or lack autonomous generation capabilities by strictly relying on pre-existing target instances. To overcome these limitations, we propose the Masked Inpainting and Residual Blending Network (MIRB-Net) for instance-free, background-preserving, multi-class coupled generation directly onto authentic measured backgrounds. Treating the background as a measured clutter carrier, MIRB-Net synthesizes class-conditional targets in situ using three synergistic mechanisms. First, a Background-Preserving Residual Editing (BPRE) mechanism explicitly decouples spatial blending from residual prediction, effectively confining generative perturbations to the target vicinity. Second, a self-supervised Masked Inpainting Pretraining (MIP) strategy specifically designed for BPRE equips the network with crucial priors of target scattering and background statistics. Third, joint global–local adversarial supervision alongside tri-region spatial constraints ensures seamless radiometric coupling across target–clutter boundaries. Evaluated under a dedicated multidimensional assessment framework on the MSTAR dataset, MIRB-Net demonstrates superior local structural fidelity, faithful background preservation, and natural target–clutter coupling. Furthermore, hybrid data-augmentation experiments confirm that the autonomously generated samples consistently improve downstream ATR performance across diverse deep learning architectures.