Hessian Matrix-Based Improved RIFT Algorithm for Infrared-Visible Image Registration

Traditional feature matching algorithms such as SIFT and SURF rely on image gradient information to construct descriptors, but their performance significantly degrades in matching visible and infrared images due to nonlinear radiometric differences(NRD). The RIFT enhances cross-modal matching robustness through phase consistency (PC) feature detection and Maximum Index Map (MIM) descriptor design. However, its core log-Gabor filter group introduces high computational complexity due to multi-scale dense convolution. To address this, this paper proposes an improved RIFT algorithm based on pre-screening acceleration: First, a fast feature point pre-screening mechanism using Hessian matrix is designed to rapidly locate potential corner and edge regions via rectangular region integral responses, reducing redundant logGabor convolution in subsequent steps. Second, by analyzing radiometric differences between infrared and visible images, parameter optimization for the MIM descriptor is conducted specifically for infrared-visible scenarios. Experimental results demonstrate that the improved algorithm reduces single-frame processing time of RIFT on infrared-visible datasets.