Cross-spectral registration of natural images with SIPCFE

Abstract

Image registration is a viable task in the field of computer vision with many applications. When images are captured under different spectrum conditions, a challenge is imposed on the task of registration. Researchers carefully handcraft a local module insensitive to illumination changes across cross-spectral image pairs to tackle this challenge. We, in this paper, develop an optimized feature-based approach Single Instance Phase Congruency Feature Extractor (SIPCFE) to tackle the problem of natural cross-spectral image registration. SIPCFE uses the phase information of an image pair to quickly identify and describe reliable keypoints that are insensitive to illumination. It then employs a sequence of outlier removal processes to find the matching feature points accurately and the Direct Linear Transformation to estimate the geometric transformation to align the image pair. We extensively study the proposed approach for every module in the system to give more insights into the challenges. We benchmark our proposed method and other state-of-the-art feature-based methods developed for cross-spectral imagery on three datasets with various settings and image contents. The comprehensive analysis of cross-spectral registration results of natural images demonstrates that SIPCFE achieves up to 47.24%, 14.29%, and 12.45% accuracy improvement on the first, second, and third dataset, respectively, over the second best registration method in the benchmark.