Regional Infrared and Visible Image Registration for Distribution Substation Areas With Wide Depth of Field

Abstract

Infrared and visible image registration is a key task in automated distribution network inspection. Previous approaches typically apply a single homography to align infrared and visible images leading to local misalignment. To address this, this study proposes a novel region-based registration algorithm specifically designed for wide-depth distribution substation areas inspection. The proposed method begins by generating matched keypoints using SuperPoint and LightGlue. Second, Adaptive label k-means is employed to assign k-category labels to the paired keypoints. Third, the region annotation module and SegmentAnything 2 propagate the point labels to region labels. Finally, different homographies are applied to the corresponding regions based on the labels with a depth rule used to determine which pixels should be preserved in overlapping areas. Experimental results show that the proposed method accurately registers infrared and visible images across multi-depth planes, achieving a root mean square error of 5.53 and a precision of 0.752, striking a balance between registration accuracy and processing speed.