Non-Rigid Registration for Infrared and Visible Images via Gaussian Weighted Shape Context and Enhanced Affine Transformation

Abstract

Image registration is a prerequisite for image fusion from multiple modalities, such as infrared (IR) and visible (VIS) images. Although there have been many various methods of image registration, non-rigid registration for IR and VIS images is still challenging due to large differences between IR and VIS images. In this work, a point feature-based method is proposed to improve the performance on non-rigid IR and VIS image registration. Firstly, a feature descriptor -Gaussian weighted shape context (GWSC) -is improved from shape context (SC) to fast extract matching point pairs from edge maps in IR and VIS images. With the set of matching point pairs, a Gaussian-field-based objective function is established to measure the distance between IR and VIS images. Then, the enhanced affine transformation (EAT) model is proposed to generalize affine model from linear to non-linear case and describe the regularity of global deformation between IR and VIS images. At last, the derivative of the distance measure is expressed with respect to the EAT model and thus, the optimal parameters are estimated by using the quasi-Newton method. The qualitative and quantitative comparisons demonstrate that the proposed method (GWSC-EAT) can be successfully applied to non-rigid registration of IR and VIS images and moreover, it is superior to the state-of-the-art methods on the accuracy and speed of non-rigid registration.

Classification of RIGID and Non-Rigid Transformations with Autoencoder Representations

“…However, feature matching on non-rigid images is not only a medical application. Another recent example of non-rigid image matching is for image registration, an important function for fusing different image sources, such as in the GWSC-EAT proposed by Min et al in 2020 [6]. These domain-specific non-rigid feature matching algorithms work very well on the problem sets they were designed for, but are not very generalizable to other problems or rigid transformations.…”

A Unified Framework for Nonrigid Point Set Registration via Coregularized Least Squares

“…The problem of registering point sets arises in a variety of applications, such as workpiece localization [1], digital elevation model generation [2], shape recognition [3], [4], and image registration [5], [6], thus laying a foundation for the implementation of many high-technology platforms. In general, the task of point set registration is to assign potential point-to-point correspondences and recover the underlying transformation that warps one point set to the other.…”