Constraint relationship for reflectional symmetry and rotational symmetry

Dinggang Shen, Horace Ho-Shing Ip, Eam Khwang TeohJ. Electronic Imaging 2001 · DOI

TL;DR

This paper introduces a novel theorem that establishes a direct relationship between the number of reflectional symmetry axes and the fold number of rotational symmetry in planar images. This theorem provides a crucial constraint for symmetry detection algorithms, allowing for verification of results and guidance in the detection process.

What Problem Does This Paper Address?

Detecting symmetries in images is a fundamental task in computer vision, with applications ranging from object recognition to image analysis. While techniques for detecting reflectional symmetry (mirror symmetry) and rotational symmetry (invariance to rotation) have been developed independently, there's been a lack of understanding regarding how these two types of symmetry relate to each other within a single image. This gap makes it challenging to cross-validate symmetry detection results from different methods and can lead to erroneous detections. The paper aims to bridge this gap by defining a precise mathematical relationship between the number of reflectional symmetry axes and the order of rotational symmetry.

Methods

The core of the paper is the development and proof of a theorem that links reflectional and rotational symmetries. The authors first define reflectional symmetry as invariance to reflection across one or more lines (reflection-symmetric axes) and rotational symmetry as invariance to rotation by $2\pi/K$ radians, where $K$ is the largest such integer (the fold number).

The key theorem, Theorem 3, states:

  1. For a $K$-fold rotationally symmetric image ($K \ge 1$), if it possesses any reflection-symmetric axes, the number of these axes ($M$) must be either 0 or $K$.
  2. Conversely, if a non-zero number of reflection-symmetric axes ($M > 0$) are detected, then $M$ must be equal to the image's fold number ($K$).

The proof involves analyzing the properties of images under both types of transformations. It demonstrates that if an image has $M$ reflection-symmetric axes, their angles must be uniformly distributed, and this uniformity, when combined with rotational symmetry, forces the number of reflection axes to be equal to the fold number.

Datasets

Not discussed in the paper. The paper uses illustrative examples and experimental results shown in figures, but does not refer to specific, named datasets.

Key Insights and Ideas

  • A Fundamental Constraint: The paper reveals a surprisingly simple yet powerful constraint: the number of reflectional symmetry axes in a rotationally symmetric image is either zero or exactly equal to its fold number. This is a significant insight that wasn't previously established.
  • Cross-Validation Tool: This theorem acts as a built-in cross-validation mechanism. If one algorithm detects, say, 3 reflection axes and another detects 2-fold rotational symmetry, the theorem immediately flags this as an inconsistency.
  • Guidance for Detection: The theorem can guide the development and application of symmetry detection algorithms. It suggests that when searching for symmetries, one should expect a specific relationship between the counts of different symmetry types.
  • Robustness Implication: The experimental results show the theorem holds even for noisy or gray-level images, suggesting its broad applicability.

Key Results

The paper's primary result is the Theorem 3 itself, which mathematically formalizes the relationship between reflectional and rotational symmetry counts.

Experimentally, the authors demonstrate the theorem's correctness using various images:

  • Figure 1: Shows examples of images with different fold numbers (1 to 4). Images that are not reflection-symmetric have zero reflection axes, regardless of their fold number. Images that are reflection-symmetric have a number of reflection axes exactly equal to their fold number.
  • Figure 2 & 3: Extend these demonstrations to gray-level and noise-corrupted images, showing the theorem's robustness.
  • Figure 4: Illustrates the practical utility of the theorem. An example is shown where a rotational symmetry detector incorrectly identifies three fold axes for an image that is actually only 1-fold symmetric. By applying the constraint theorem and analyzing the detected reflectional symmetry (one axis), it becomes evident that the rotational symmetry detection is erroneous.

Limitations and Future Work

  • Planar Images: The theorem is explicitly stated for "planar images." Its direct applicability to 3D objects or projections might require further investigation.
  • Ideal Symmetry: The proofs and theorems assume perfect symmetry. The paper touches upon "almost symmetric" images and robustness to noise, but the exact behavior and limitations for images with significant deviations from perfect symmetry are not exhaustively explored.
  • Algorithm Integration: While the paper demonstrates how the theorem can be used to guide symmetry detection (as shown in Figure 4), it doesn't propose a new, unified algorithm that integrates this constraint into the detection process itself. Future work could focus on developing such algorithms.
  • Generalization: The paper focuses on the number of axes. Exploring how the orientations of reflectional axes relate to the orientations of rotational symmetry in more complex scenarios could be a future direction.

Summary updated

BibTeX
@article{shen2001constraint,
  author = {Dinggang Shen and Horace Ho-Shing Ip and Eam Khwang Teoh},
  title = {{Constraint relationship for reflectional symmetry and rotational symmetry}},
  journal = {J. Electronic Imaging},
  year = {2001},
  doi = {10.1117/1.1407264},
}
Figure