Robust Optical-to-SAR Image Registration via Dense Tukey-Weighted Gradient Histogram and Structural Saliency Weight

Abstract

Optical-to-SAR image registration is a fundamental prerequisite for multisource remote sensing applications, yet it remains challenging due to severe nonlinear radiometric differences, complex speckle noise, and structural blurring. Existing area-based matching methods often impose uniform spatial weighting. This uniformity makes them highly susceptible to mismatches in textureless or noise-dominated regions. To address this issue, we propose a robust cross-modal matching framework driven by a novel dense Tukey-weighted gradient histogram (DTGH) descriptor and a structural saliency weight (SSW). Specifically, DTGH employs directional soft assignment and spatial aggregation via a 2-D Tukey window to extract modality-invariant geometric structures. This approach effectively mitigates quantization boundary artifacts and prevents feature blurring. Furthermore, the SSW introduces a spatial confidence measure derived from local variance statistics. It adaptively shifts the matching focus toward highly reliable structures while suppressing the interference of homogeneous areas. Crucially, the SSW serves as a lightweight, plug-and-play module that can be seamlessly integrated to enhance existing multimodal registration pipelines. Comprehensive evaluations across patch-level, full-image, and large-scale scenarios demonstrate that the proposed method achieves state-of-the-art accuracy and reliability. It consistently outperforms both representative hand-crafted baselines and advanced deep learning models, while circumventing the domain shift bottlenecks inherent in data-driven approaches. Consequently, this training-free framework exhibits exceptional cross-domain generalization and high computational efficiency in CPU-only environments.