A Method of Generating Calligraphy of Japanese Character using Deformable Contours

IJCAI 1997

TL;DR

This paper introduces a novel method for generating diverse Japanese calligraphy styles from a limited set of existing fonts. By modeling character strokes as deformable contours (g-snakes) and applying energy minimization principles, the system learns global and local stroke deformations to create new, visually appealing calligraphy.

What Problem Does This Paper Address?

The generation of diverse calligraphy styles for characters, particularly in languages like Japanese with complex stroke structures, is a challenging problem. Existing methods are often limited to a predefined set of fonts or rely on complex, manually defined rules that lack generality. The authors aim to bridge this gap by developing a system that can automatically generate a range of calligraphy styles from just two sample fonts, with minimal human intervention. This addresses the need for flexible and automated calligraphy generation that captures the nuances of human writing.

Methods

The core of the proposed method is the g-snake model, a type of deformable contour. Each stroke of a character is represented as a g-snake, which is essentially a chain of points (snaxels). The generation process involves two main types of deformation:

  • Global Deformation: This accounts for rigid motions like scaling, rotation, and stretching between corresponding strokes of the two sample fonts. The paper leverages an invariance property of the g-snake representation to compute these transformations efficiently using affine transformations.
  • Local Deformation: This is modeled using the energy minimization principle inherent in the g-snake. The g-snake's internal energy keeps its shape, while an external force, derived from the difference between the initial and target stroke positions, pulls it towards a desired configuration. The paper formulates this as a Markov Random Field problem, where the probability of a contour is defined by its internal and external energies.

The system learns the parameters for these deformations by treating the problem as a Maximum Posterior (MAP) estimation. Specifically, it learns local regularization parameters that control the degree of local stroke deformation. By varying a control parameter t (ranging from 0 to 1), the system can interpolate between the two input calligraphy styles, generating a spectrum of intermediate styles.

Datasets

The paper uses existing Japanese Hiragana fonts (kaisho, gyosho, mini, gothic) as sample calligraphy for training and evaluation. They also tested the system on a real scanned calligraphy of Hiragana "a". No new datasets were introduced.

Key Insights and Ideas

  • Deformable Contours for Calligraphy: The g-snake model is a powerful tool for representing and deforming character strokes, capturing both global and local shape changes essential for calligraphy.
  • Energy Minimization as a Generative Principle: Framing calligraphy generation as an energy minimization problem, where strokes are pulled towards a target shape, yields natural-looking results.
  • Controlling Deformation with a Single Parameter: The use of a single parameter t to control the interpolation between two styles provides an intuitive and effective way to generate a range of calligraphic variations.
  • Bridging the Gap Between Fonts: The method successfully interpolates between distinct calligraphy styles, demonstrating its ability to generate intermediate forms that are visually plausible.

Key Results

The paper demonstrates the effectiveness of its method through several examples of generated Hiragana calligraphy.

  • Visual Interpolation: Figures 6 and 7 show generated calligraphy for Hiragana "a", "i", "u", "e", and "o", illustrating the smooth transition from one font style to another as the control parameter t increases.
  • Real Calligraphy Integration: Figure 8 shows generated calligraphy based on a real scanned Hiragana "a" and a "kaisho" font, indicating the system's applicability to real-world data.
  • Qualitative Assessment: The authors state that the generated calligraphy shows a "reasonable way of generating characteristic calligraphy" and that the results are "similar to calligraphy written by real people."

Limitations and Future Work

The authors acknowledge several limitations:

  • Artistic Acceptability: Some generated results are not considered artistically perfect, and as t increases, the character can lose the rhythm of the initial character or become non-canonical.
  • Local Blurring: Due to the energy minimization process and control difficulties, some strokes in the generated characters can be locally blurred.

Future work suggested includes:

  • Automatic Stroke Extraction: Developing methods to automatically extract strokes from characters.
  • Modeling Writing Dynamics: Incorporating models for brush force, pressure, and writing dynamics.
  • Interpolating Between Strokes: Generating semi-cursive styles by interpolating between strokes.
  • Extension to Chinese Characters: Adapting the system for the more complex calligraphy of Chinese characters.

Summary updated

BibTeX
@inproceedings{1997method,
  title = {{A Method of Generating Calligraphy of Japanese Character using Deformable Contours}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1997},
}
Figure