ISAR Cross-Range Scaling via Siamese Network with Multiscale Convolutional Attention

Inverse synthetic aperture radar (ISAR) image cross-range scaling is a prerequisite for accurate target information interpretation. High-precision cross-range scaling relies on the accurate estimation of the target's rotation parameters, which typically involves feature extraction and correlation processing between adjacent images. However, ISAR images exhibit strong sparsity and significant anisotropy, leading to poor robustness and time-consuming challenges in image feature extraction and correlation, thereby hindering the demand for robust and rapid ISAR image scaling. To address this issue, based on the analysis of the relationship between cross-range scaling and rotation features in adjacent-frame ISAR images, this paper proposes a cross-range scaling method using a Siamese network with multiscale convolutional attention (MCA-SN). The method employs the Siamese network (SN) architecture to establish a unified feature extraction framework for adjacent images. By integrating multiscale convolution and attention mechanisms, it effectively extracts rotation-related discriminative features between adjacent images, achieving end-to-end mapping from images to rotation parameters. Simulation experiments demonstrate that the proposed method significantly improves the estimation accuracy and timeliness of rotation parameters, enabling robust and rapid high-precision ISAR image scaling.