Structure-Preserving Neural Style Transfer.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Current neural style transfer methods use high-level features from classification networks, which can neglect details and distort content structures.
• These methods often focus on central objects, leading to scattered style textures and disrupted spatial distributions in stylized outputs.

Purpose of the Study:

• To develop a novel image stylization method that addresses the limitations of existing approaches by incorporating an explicit structure representation.
• To enhance the preservation of content structures and improve the overall visual effectiveness of stylized images, particularly for images sensitive to distortion.

Main Methods:

• The proposed method introduces an additional structure representation alongside traditional style and content representations.
• Structure representation combines global structure (depth map) and local structure details (image edges) to capture spatial distributions and object structures.
• Leverages pretrained classification networks for feature extraction while augmenting with depth and edge information.

Main Results:

• Experimental results demonstrate impressive visual effectiveness of the proposed method.
• The method significantly improves stylization for images sensitive to structure distortion, such as those with multiple objects or clear dominant structures.
• Preserves content structures more effectively compared to methods relying solely on high-level features.

Conclusions:

• The novel structure representation is effective in improving neural style transfer.
• The method offers a significant advancement for stylizing images where structural integrity is crucial.
• This approach provides a more robust solution for complex image stylization tasks.