PSA-HWT: handwritten font generation based on pyramid squeeze attention

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing handwriting font generation models, often based on Convolutional Neural Networks (CNNs) and Transformers, struggle with detailed feature extraction, leading to suboptimal font quality.
• Specific issues include insufficient extraction of overall font structure, stroke thickness, and stroke curvature, limiting the realism and detail of generated fonts.

Purpose of the Study:

• To propose a novel handwritten font generation model, PSA-HWT, that addresses the limitations of current methods by incorporating Pyramid Squeeze Attention (PSA).
• To enhance the extraction of multi-scale spatial information and improve the understanding of fine-grained font features like shape, thickness, and curvature.
• To ensure the coherence and quality of generated handwritten text through an effective encoder-decoder architecture with self-attention mechanisms.

Main Methods:

• The proposed PSA-HWT model utilizes a two-part encoder-decoder architecture.
• The encoder employs a multi-branch structure for multi-scale feature extraction, capturing semantic information and global font structure.
• The decoder incorporates a self-attention mechanism to model dependencies within the generated sequence, ensuring stroke and character coherence.

Main Results:

• Experiments on the IAM dataset show PSA-HWT achieves a 16.35% reduction in Fréchet Inception Distance (FID) score.
• The model also demonstrates a 13.09% decrease in Geometry Score (GS) compared to state-of-the-art methods.
• These results indicate a significant improvement in the quality and detail of generated handwritten fonts.

Conclusions:

• The PSA-HWT model effectively addresses limitations in feature extraction for handwritten font generation.
• The integration of Pyramid Squeeze Attention and a self-attention mechanism leads to superior font quality and practical value.
• PSA-HWT represents a significant advancement in generating realistic and high-fidelity handwritten fonts.