School of Mathematical & Statistical Sciences Faculty Publications

Multimodal Prior-Guided Generative Inversion for Robust ISAR Imaging

Document Type

Article

Publication Date

2026

Abstract

Inverse Synthetic Aperture Radar (ISAR) imaging is a classical electromagnetic inverse problem, where the reconstruction process is inherently ill-posed due to insufficient observational data and noise interference. Traditional methods rely on mathematical priors (such as sparsity) or priors learned from quantitative data, but struggle to effectively utilize semantic-level multimodal prior information. This paper proposes a Multimodal Prior-Guided ISAR (MPGI) framework that jointly uses observed radar measurements, textual descriptions, and contour priors for inverse ISAR reconstruction. The MPGI framework comprises: (1) A frozen pre-trained BERT text encoder for encoding high-level textual priors; (2) A Physical Adapter Model for channel-wise concatenating and encoding ISAR measurements and spatial contour priors; (3) A Multimodal Generative Model (MGM) for fusing multimodal embeddings and synthesizing the final image. The physical adapter and generative model employ are jointly optimized during training. Experiments on simulated datasets and measured Yak-42 radar data demonstrate that MPGI achieves improved reconstruction quality under the evaluated noise and phase-error conditions. Furthermore, the model’s reliance on contour priors indicates that the physical adapter has learned meaningful and non-trivial spatial structural dependencies. This work provides a flexible and efficient solution for leveraging multimodal semantic priors in radar imaging, with broad application prospects in intelligent sensing and next-generation wireless communication systems.

Comments

Copyright © 2026, IEEE

Publication Title

IEEE Internet of Things Journal

DOI

10.1109/JIOT.2026.3740398

Link to Full Text

Share

COinS