A Novel Shape Representation Method for Complex Trademark Image

Ben Ye, Zhanchuan Cai, Ting Lan, Youqing XiaoIEEE Access 2019 · DOI

Introduction

Representing trademark images in computer vision is a significant challenge due to their inherent complexity. Modern logos often consist of multiple disjoint shapes, text, and intricate geometric patterns. Traditional mathematical tools like Fourier series or continuous wavelets frequently struggle with the sharp "discontinuities" at the edges of these shapes. This leads to the Gibbs phenomenon, where reconstructions exhibit "ringing" artifacts or oscillations at the boundaries. To achieve high fidelity, these traditional methods require a large number of terms, making them computationally inefficient for large-scale retrieval systems.

Methodology: The Hierarchical V-system

To address these limitations, this research introduces the Hierarchical V-system (HV-system). This framework utilizes piecewise orthogonal functions to represent complex shapes with high precision and minimal data terms.

Flowchart of the Hierarchical V-system (HV-system) for shape representation.

The HV-system operates through three primary mechanisms:

  1. Multi-Degree Hierarchy: The system is constructed using multi-degree $k$ polynomials (where $k = 0, 1, 2, 3$). While degree-0 systems (Haar wavelets) capture basic structures, higher degrees utilize Legendre polynomials to refine smooth details.
  2. Iterative Refinement: The algorithm begins with a base representation and calculates the error (the "norm") between the original shape and the reconstruction. If the error exceeds a threshold, additional layers of the hierarchy are added to "fill in" missing details.
  3. Invariant Descriptors: The authors derive HV descriptors that are mathematically proven to be invariant to rotation, translation, and scaling, ensuring robust recognition regardless of the logo's orientation or size.

Dataset and Evaluation

The performance of the HV-system was evaluated against standard benchmarks and a specialized dataset of complex trademarks.

The custom dataset of 32 complex trademark images used for testing reconstruction efficiency.

The evaluation included:

  • MPEG-7 Shape Database: A standard benchmark with 1,400 images.
  • Kimia’s Database: 99 images featuring partially occluded shapes.
  • Custom Trademark Set: 32 binary complex trademark images specifically chosen to test multi-object shape reconstruction.

Key Results

The HV-system demonstrates significant improvements in both efficiency and visual fidelity:

  • Term Reduction: For a hexagonal logo, the HV-system required only 311 terms to reach a precision that required 765 terms for a Fourier series and 558 for a Haar wavelet.
  • Visual Fidelity: Unlike Fourier methods that produce "fuzzy" edges, the HV-system maintains sharp, faithful reconstructions of multi-object geometry.
  • Robust Invariance: Experimental results confirmed that the "energy" frequency descriptors remain stable even under significant geometric transformations like rotation and scaling.

Conclusion

The Hierarchical V-system offers a powerful alternative for trademark representation by leveraging the strengths of piecewise polynomials. By only applying complex mathematical terms where the shape is actually complex, the system achieves a "sparse" and efficient representation. Future work will focus on automating the error threshold selection and integrating these descriptors into large-scale industrial trademark retrieval engines.

Summary updated

BibTeX
@article{ye2019novel,
  author = {Ben Ye and Zhanchuan Cai and Ting Lan and Youqing Xiao},
  title = {{A Novel Shape Representation Method for Complex Trademark Image}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2912230},
}
Figure