Arbitrary Style Transfer with Multiple Self-Attention

Abstract

Style transfer aims to transfer the style information of a given style image to the other images, but most existing methods cannot transfer the texture details in style images well while maintaining the content structure. This paper proposes a novel arbitrary style transfer network that achieves arbitrary style transfer with more local style details through the cross-attention mechanism in visual transforms. The network uses a pre-trained VGG network to extract content and style features. The self-attention-based content and style enhancement module is utilized to enhance content and style feature representation. The transformer-based style cross-attention module is utilized to learn the relationship between content features and style features to transfer appropriate styles at each position of the content feature map and achieve style transfer with local details. Extensive experiments show that the proposed arbitrary style transfer network can generate high-quality stylized images with better visual quality.