Multi-orientation multi-scale aggregated registration with cyclic back-projection for multimodal images

High-precision multimodal registration is essential for multimodal image analysis and visual perception systems, yet remains challenging due to severe non-linear radiometric distortions and geometric variations. However, existing methods often fail to obtain an effective and modality-independent structural representation, resulting in poor matching performance in such heterogeneous scenarios. To address this, we propose a robust framework named MMAR. First, a Multi-Oriented Multi-Scale Gradient (MMG) space is constructed to generate a structural basis resistant to radiometric distortions. Second, a Hierarchical Sparsification of Texture-Enhanced Features (HSTF) strategy, coupled with Spatial-Similarity Non-Maximum Suppression (SS-NMS), is introduced to optimize the trade-off between feature abundance and distinctiveness. Third, a Cyclic Iterative Backward Projection (CIBP) mechanism is integrated to progressively compensate for large geometric deviations. Validated on four heterogeneous datasets covering diverse terrains, MMAR significantly outperforms ten state-of-the-art baselines. Notably, the proposed method maintains stable registration performance under significant geometric transformations, including rotations from − 90 ° to 120° and scale variations from 0 . 4 × to 4 . 5 × . These results confirm MMAR as a reliable registration foundation for practical applications in remote sensing, medical imaging, and cross-modal visual understanding.