Design and Implementation of Dongba Character Font Style Transfer Model Based on AFGAN

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Humanities

Background:

• Dongba characters, an ancient ideographic script, possess unique abstract expressions distinct from modern Chinese characters.
• Existing font style transfer methods are inadequate for Dongba characters due to their unique script characteristics.

Purpose of the Study:

• To develop an effective font style transfer method for Dongba characters.
• To adapt existing script styles (e.g., small seal, slender gold) to Dongba characters using a novel Generative Adversarial Network (GAN).

Main Methods:

• Proposed an Attention-based Font style transfer Generative Adversarial Network (AFGAN) tailored for Dongba character images.
• Incorporated void constraint and font stroke constraint modules within the AFGAN architecture.
• Integrated the Convolutional Block Attention Module (CBAM) to enhance feature learning and adapt to diverse input styles.

Main Results:

• The AFGAN method demonstrated superior performance in evaluation metrics and visual quality compared to existing networks.
• Successful style transfer of small seal script and slender gold script to Dongba characters was achieved.
• Quantitative and qualitative analyses, including expert artist evaluation, validated the effectiveness of the proposed method.

Conclusions:

• The AFGAN method is effective for transferring font styles to Dongba characters.
• The approach successfully captures and applies style features from different scripts, preserving the integrity of Dongba characters.