MTCPNet: A Mamba-Based Registration Network with Tri-Branch Consistency Projection for SAR-Visible Image Registration

Abstract

Visible and synthetic aperture radar (SAR) images exhibit substantial nonlinear radiometric differences and geometric deformations due to their fundamentally different imaging mechanisms, making high-precision registration between the two modalities a long-standing challenge in remote sensing image processing. Existing deep learning-based cross-modal registration methods mostly adopt purely convolutional architectures or hybrid convolution–Transformer frameworks, which struggle to achieve a favorable trade-off between long-range dependency modeling and computational efficiency. Moreover, current methods generally rely only on heterogeneous cross-modal supervision for end-to-end training, while overlooking the geometric deformation priors embedded in intra-modal consistency. To address these issues, this paper proposes MTCPNet, a Mamba-based registration network with tri-branch consistency projection for SAR–visible image registration. Specifically, a feature consistency projection module is designed to project SAR and visible images into a modality-invariant shared feature space, with a feature consistency loss introduced to explicitly constrain cross-modal geometric alignment. A Mamba-based hybrid architecture serves as the feature extraction backbone, integrating the linear-complexity long-range dependency modeling of selective state space models with the local contextual representation of window-based self-attention. Under a tri-branch training paradigm, intra-modal consistency supervision and cross-modal matching supervision are jointly incorporated to optimize network parameters. Experimental results demonstrate that MTCPNet consistently outperforms state-of-the-art methods on multiple benchmark datasets, providing a promising solution for high-precision multisource remote sensing image registration.