Efficient preservation of detail features for image arbitrary style transfer
Hengchang Jiang, Duzhen Zhang, Tao Wang · 2023
Some mainstream image arbitrary style transfer models still have limitations in maintaining the saliency information and details of content images, and the generated images often have issues such as content blur and distorted distortion. To solve the problems, the style transfer model proposed in the article can preserve the details of the content image effectively. Our model includes introducing a positional self-attention module which can enhance the representation of extracted image features and effectively capture remote relationships and coordinate global and local features. The feature fusion module guided by global information is proposed, which can high-quality fuse content features and style features. A detail preservation module is proposed to preserve details. In addition, we proposed a new loss function, which can preserve the global structure of images and eliminate artifacts. The experimental results demonstrate the effectiveness of the proposed image arbitrary style transfer model in achieving a balance between style and content. The model effectively preserves the complete semantic information and detailed features of the content images, resulting in stylized images that are visually more appealing.