Abstract
The proposed algorithm detects globally the symmetry axes inside an image plane. The main steps are as follows: We firstly extract edge features using Log-Gabor filters with different scales and orientations. Afterwards, we use the edge characteristics associated with the textural and color information as symmetrical weights for voting triangulation. In the end, we construct a polar-based voting histogram based on the accumulation of the symmetry contribution (local texture and color information), in order to find the maximum peaks presenting as candidates of the primary symmetry axes.
Introduction
Symmetry is a fundamental feature in visual perception and a critical cue for object recognition and scene understanding. However, detecting reflection symmetry in "natural" images remains a significant challenge due to complex backgrounds, varying lighting conditions, and intricate textures that can confuse traditional algorithms.
Most existing methods rely on standard Gabor filters or simple edge gradients to identify symmetrical points. These approaches often struggle with the "DC component"—the average brightness of an image—and typically ignore the rich discriminative information provided by color. This paper introduces a robust algorithm that addresses these limitations by combining Log-Gabor wavelet features with textural and color histograms to create a more resilient symmetry detection framework.
Methodology
The proposed pipeline refines the symmetry detection process through a multi-step approach focused on high-fidelity feature extraction and weighted voting.
1. Log-Gabor Edge Detection and Feature Extraction
Instead of standard Gabor filters, the algorithm utilizes Log-Gabor filters. These are superior because they have no DC component, meaning they are not biased by overall image brightness, and they can represent edge details more sharply across multiple scales and orientations. The image is divided into a grid, and in each cell, a feature point is extracted based on the strongest wavelet response.
2. Dual Histograms: Texture and Color
For every extracted feature point, the algorithm computes two types of descriptors to ensure high discriminative power:
- Textural Histograms: These capture local edge orientations.
- Color Histograms: These capture the distribution of Hue, Saturation, and Value (HSV), providing a critical layer of information often ignored in grayscale-based methods.
3. Symmetry Triangulation and Voting
This is the core of the algorithm. Every pair of feature points "votes" for a potential symmetry axis. The weight of this vote is determined by the product of three similarities: orientation similarity (geometric mirroring), textural similarity (pattern consistency), and color similarity (chromatic consistency). These votes are accumulated in a 2D polar histogram representing the axis angle and its distance from the origin. The highest peaks in this histogram correspond to the primary symmetry axes.
Experimental Evaluation
The authors evaluated their method against three major benchmarks: the PSU Datasets (single and multiple symmetry), the AVA Dataset (aesthetic-based), and the NY Datasets (stable ground truth).
The proposed color-enabled version (LgC) consistently outperformed state-of-the-art algorithms such as Loy2006, Cic2014, and Ela2016. For instance, in the PSU single symmetry dataset, LgC identified 116 correct symmetries, a significant improvement over the 97 identified by the previous best method (Ela2016). Precision-Recall curves further demonstrate that LgC maintains higher precision at higher recall levels, particularly in complex "Multiple Symmetry" scenarios.
Key Insights and Conclusion
A primary insight of this research is the effectiveness of Log-Gabor filters in producing "cleaner" amplitude maps by suppressing DC component noise. Furthermore, the integration of color serves as a powerful pruning tool; while many points in a complex scene may share similar textures by chance, they are far less likely to share identical HSV distributions unless they are truly symmetrical.
While the algorithm represents a significant leap in performance, future work will address the quadratic computational complexity of the voting process ($P(P-1)/2$) and refine peak selection to reduce false positives in highly textured backgrounds. Extending this robust wavelet-histogram approach to rotational and translational symmetry remains a promising avenue for future research.
Summary updated