Spherical Binary Images Matching

Wei Liu, Yuanjun HeLecture Notes in Computer Science 2007 · DOI · OpenAlex

Abstract

Matching patterns on a spherical surface is a complex task due to the lack of natural boundaries or anchor points. This paper introduces a framework for matching spherical binary images (SBIs) by calculating their maximal superposition degree. By discretizing the sphere using an icosahedron-based geodesic grid and employing a two-step rotation strategy, the authors provide a robust method for comparing spherical data. Results show that moderate subdivision levels ($N=3$) provide the best balance between detail and robustness.

Introduction

Traditional image matching techniques rely on planar coordinate systems with clear borders and corners. However, spherical surfaces are finite yet unbounded, making alignment difficult. In fields such as remote sensing and 3D model retrieval, data is frequently mapped to spheres. Determining if two binary patterns are identical despite different rotations requires a finite search space for the "perfect rotation." This paper proposes a method to create such a space and accurately measure similarity.

Methodology

The proposed algorithm follows a multi-step process to discretize the sphere and identify optimal alignment:

  1. Icosahedral Discretization: A regular icosahedron serves as the base to ensure uniform division (isotropy). Its 20 triangular facets are projected onto a unit sphere.
  2. Geodesic Subdivision ($N$): Each facet is subdivided into smaller trigonal grids $N$ times. For $N=3$, the sphere comprises 1,280 cells, each tagged as 1 or 0 based on the underlying image data.
Geodesic subdivision levels N=2, 3, 4 on a sphere.
  1. Two-Step Rotation Strategy:
    • Intra-Group Rotation: The algorithm checks 60 specific rotations corresponding to the icosahedron's Symmetrical Group.
  2. Inter-Group Relaxation ($L$): To account for random initial orientations, the authors use a relaxation mechanism to generate $L$ uniform starting orientations across the sphere.
Examples of spherical binary patterns under different rotations.
  1. Similarity Metric: Similarity ($S$) is defined as the ratio of matching tags between two images at a specific orientation. The final result is the maximum $S$ found across all $60 \times L$ rotations.

Analysis and Results

The authors evaluated 18 combinations of subdivision levels ($N$) and rotation groups ($L$). The optimal configuration was found at $N=3$ and $L=15$, yielding 37.9% accuracy.

A key finding is the "Resolution Paradox": while higher resolution usually implies better matching, $N=3$ outperformed $N=4$. Excessive detail at $N=4$ led to an "explosion of information," making the general shape description less robust—a form of spatial overfitting. Additionally, the authors established a formula to estimate maximum angular error ($A$) based on $L$, allowing for a trade-off between computational speed and precision.

Conclusion

While the current method achieves a modest success rate, it represents a significant step toward structured spherical image retrieval. The reliance on brute-force superposition checks presents a computational bottleneck. Future research will focus on developing global feature descriptors, similar to SIFT for planar images, to enable faster off-line matching in large-scale databases.

Summary updated

BibTeX
@inproceedings{liu2007spherical,
  author = {Wei Liu and Yuanjun He},
  title = {{Spherical Binary Images Matching}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2007},
  doi = {10.1007/978-3-540-72586-2\_19},
}
Figure