Dual-Affinity Style Embedding Network for Semantic-Aligned Image Style Transfer.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Image style transfer synthesizes images using content from one source and style from another.
• User perception of style transfer quality is significantly influenced by semantic correspondence between content and style.
• Existing methods often transfer global style without semantic alignment, and current semantic approaches are computationally expensive due to iterative optimization.

Purpose of the Study:

• To develop a computationally efficient method for semantic style transfer that aligns style at a granular, semantic region level.
• To improve the subjective quality of stylized images by ensuring semantic correspondence.
• To address the limitations of existing methods regarding computational cost and semantic alignment.

Main Methods:

• Introduced a novel dual-affinity style embedding network (DaseNet) for semantic style transfer.
• Developed a dual-affinity module to jointly model feature correlation and semantic correspondence for embedding local style patterns.
• Incorporated semantic-weighted style loss and region-consistency loss to ensure semantic alignment and content preservation.

Main Results:

• DaseNet synthesizes images with style aligned at semantic region granularity.
• The proposed method effectively balances visual quality and inference efficiency.
• Experimental results demonstrate the effectiveness of DaseNet across different scene categories.

Conclusions:

• DaseNet offers an efficient and effective solution for semantic style transfer.
• The dual-affinity module and novel loss functions enable precise semantic alignment.
• The method advances the state-of-the-art in image style transfer by focusing on semantic correspondence.