Image style transfer with collection representation space and semantic-guided reconstruction

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Art

Background:

• Current image style transfer methods struggle with capturing the full scope of an artist's style from a single image.
• Previous approaches often compromise content preservation, leading to poor structural integrity and reduced comprehensibility in stylized images.

Purpose of the Study:

• To address limitations in style resemblance and content preservation in image style transfer.
• To improve the visual quality and comprehensibility of stylized images.

Main Methods:

• Developed a style transfer system utilizing a collection representation space and semantic-guided reconstruction.
• Trained an encoder-decoder network on art collections to create a representation space reflecting an artist's style.
• Employed semantic information to guide the reconstruction of the input image's representation for enhanced content preservation.

Main Results:

• The proposed method effectively balances capturing artistic characteristics with preserving content information.
• Quantitative analysis and qualitative evaluations confirm the approach's efficacy.
• Achieved superior style resemblance and content preservation compared to existing methods.

Conclusions:

• The novel approach significantly enhances image style transfer by addressing key limitations.
• The system offers a robust solution for generating high-quality stylized images with improved content integrity.
• This work advances the field of style transfer by effectively integrating artistic style and content preservation.