Heterogeneous image registration method combining infrared and visible light

Abstract

This application proposes a heterogenous image registration method integrating infrared and visible light, aiming to address issues such as inconsistent grayscale distribution, misaligned edge structures, challenging feature matching, and difficulties in adapting to non-rigid deformations caused by differences in imaging mechanisms. The method includes: acquiring synchronized and distortion-corrected infrared and visible light images; extracting multi-scale features through a dual-stream convolutional network with shared underlying parameters; generating context-aware cross-modal matching features via a bidirectional cross-attention module; estimating local non-rigid deformation fields based on an 8×8 control grid spatial transformation network; achieving end-to-end registration through differentiable resampling; and jointly optimizing via a composite loss function comprising weighted pixel error, structural similarity, and feature consistency. This application enables high-precision, real-time cross-modal image alignment in complex scenarios, significantly enhancing registration robustness and the performance of subsequent fusion tasks.