MSEmbGAN: Multi-Stitch Embroidery Synthesis via Region-Aware Texture Generation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional neural networks (CNNs) are limited in synthesizing diverse embroidery features due to challenges in predicting varied stitch types.
• Effective extraction of stitch features is crucial for advanced embroidery synthesis from images.

Purpose of the Study:

• To propose a novel Multi-Stitch Embroidery Generative Adversarial Network (MSEmbGAN) for predicting diverse embroidery features from images.
• To develop a region-aware texture generation sub-network capable of predicting varied stitch types and textures.
• To introduce a colorization network for ensuring full image color consistency between input and output.

Main Methods:

• Developed a region-aware texture generation sub-network that classifies image regions and generates stitch textures based on shape features.
• Proposed a colorization network with a color feature extractor to maintain color consistency.
• Created and annotated a new large-scale multi-stitch embroidery dataset with over 30K images.

Main Results:

• The MSEmbGAN successfully predicts diverse embroidery features and stitch types, outperforming existing methods.
• Quantitative and qualitative evaluations, including a user study, demonstrate superior performance in embroidery synthesis and style transfer.
• The proposed method achieves state-of-the-art results across all evaluation indicators.

Conclusions:

• The MSEmbGAN represents a significant advancement in generating diverse and high-quality embroidery from images.
• The new dataset and proposed network architecture address key limitations in current embroidery synthesis research.
• This work paves the way for more sophisticated AI-driven embroidery design and creation tools.