Phase-Guided Cross-Frequency Integration Network for ISAR and Optical Image Fusion

Inverse synthetic aperture radar (ISAR) and optical image fusion aims to generate a composite image that simultaneously emphasizes the prominent contours of spacecraft from optical images and preserves the rich texture information inherent in ISAR images. However, the limited receptive fields of spatial-domain methods restrict their ability to capture global contextual dependencies among strong scattering points in ISAR images and to effectively integrate complementary optical features. To tackle this challenge, we propose a phase-guided cross-frequency integration module (PGCFIM), which exploits the intrinsic global modeling capability of the frequency domain and the semantic expressiveness of the phase spectrum. Specifically, a deep Fourier transform is employed to establish an image-wide receptive field for intra-domain global modeling. Subsequently, phase components are explicitly aggregated, and a gating mechanism is introduced to guide the integration of inter-domain long-range dependencies, enabling effective learning of complementary cross-modal representations. To eliminate reliance on hand-crafted fusion strategies, we design an end-to-end network, named PGCFINet. By jointly enhancing cross-domain interaction, frequency-domain global awareness, and explicit complementary feature integration, PGCFINet significantly strengthens cross-domain and cross-modal information interaction representation. Furthermore, to mitigate the current lack of ISAR and optical image datasets, we construct a new dataset comprising various spacecraft models, offering an alternative benchmark for evaluation. Extensive experiments demonstrate show that PGCFINet achieves superior performance than state-of-the-art methods in both qualitative and quantitative assessments. Moreover, PGCFINet is extended to infrared and visible image fusion, and the favorable results further validate its robust generalization ability. The codes of our fusion method and the dataset are forthcoming at https://github.com/WangZe0622/PGCFINet.