A Synergistic Framework Combining Multiscale Gradient Fusion and Frequency Domain Acceleration for SAR-Optical Image Registration
Lipeng Lian, Leping Chen, Daoxiang An · 2025
To address the regestration challenges stemming from inherent disparities in imaging mechanisms between synthetic aperture radar (SAR) and optical images, this study proposes this paper proposes a novel registration framework combining multiscale gradient fusion and frequency domain acceleration. The framework constructs multiscale gradient feature representations via multilevel dilated convolutions and implements an orientation adaptive allocation mechanism for gradient magnitudes using an annular Gaussian weighting strategy. This process establishes a cross modal structural correlation model in the feature space, effectively mitigating interference from SAR speckle noise and optical radiation discrepancies during feature matching. Furthermore, by transferring the traditional spatial domain optimization process to the frequency domain through fast Fourier transform (FFT), frequency domain acceleration of cross correlation computation is achieved, significantly reducing algorithmic complexity while preserving subpixel registration accuracy. Multi-scenario comparative experiments demonstrate that the proposed framework exhibits superior robustness in SAR-optical data registration tasks compared to existing image structure based methods, providing a reliable solution for multisource remote sensing collaborative interpretation.