Extrapolate azimuth angles: Text and edge guided ISAR image generation based on foundation model

Abstract

Inverse Synthetic Aperture Radar (ISAR) has been widely applied in remote sensing and space target monitoring. Automatic Target Recognition (ATR) based on ISAR imagery plays a critical role in target interpretation and pose estimation. With the growing adoption of intelligent methods in the ATR domain, the quantity and quality of ISAR data have become decisive factors influencing algorithm performance. However, due to the complexity of ISAR imaging algorithms and the high cost of data acquisition, high-quality ISAR image datasets remain extremely scarce. As a result, learning the underlying characteristics of existing ISAR data to generate large-scale usable samples has become a pressing research focus. Although some preliminary studies have explored ISAR image data augmentation, most of them rely on image sequence interpolation or conditional generation, both of which exhibit critical limitations: the former requires densely sampled image sequences with small angular intervals, while the latter can only model the mapping between limited azimuth conditions and ISAR images. Neither approach is capable of generating images of new targets under unseen azimuth conditions, resulting in poor generalization and leaving substantial room for further exploration. To address these limitations, we formally define a novel research problem, termed ISAR azimuth angle extrapolation. This task fundamentally involves high-dimensional, structured, cross-view image synthesis, requiring the restoration of visual details while ensuring physical consistency and structural stability. To address this problem, we propose ISAR-ExtraNet, a foundation-model-based framework for ISAR azimuth angle extrapolation. ISAR-ExtraNet leverages the strong representation, modeling, and generalization capabilities of pretrained foundation models to generate ISAR images of new targets under novel azimuth conditions. Specifically, the model employs a two-stage coarse-to-fine fine-tuning strategy, incorporating optical image contours and scattering center distribution constraints to guide the generation process. This design enhances both semantic alignment and structural fidelity in the generated ISAR images. Comprehensive experiments demonstrate that ISAR-ExtraNet significantly outperforms baseline methods and fine-tuned foundation models, achieving 28.76 dB in PSNR and 0.80 in SSIM. We hope that the training paradigm introduced in ISAR-ExtraNet will inspire further exploration of the ISAR azimuth extrapolation problem and foster progress in this emerging research area.