HAST: A New Style Transfer Network Integrating Convolution and Attention Mechanism

Style transfer is a computer vision technique that aims to apply the artistic style of one image to the content of another, creating a new image that preserves the content of the original image while incorporating the desired artistic style. However, existing style transfer networks still face issues with unclear semantic representation and insufficient detail preservation in stylized images. To address these problems, this paper proposes a novel style transfer network called HAST, which combines convolution and attention mechanisms. Convolution operations help preserve the detailed features, content structure, and semantic information of the image, while the attention mechanism allows the network to focus on important regions or features during image processing, resulting in stylized images with clear details and complete semantics. In the HAST model, the CPCA attention is first improved, enabling the enhanced attention to better focus on image details and adaptively adjust weights according to the network's requirements. Additionally, the image feature extractor SCCA module is designed by combining Parc convolution and the improved CPCA attention, which fully extracts semantic information from the content image and style features from the style image, preparing for subsequent feature fusion. Experimental results show that, with the above design, the HAST network generates images that not only achieve better stylization but also retain clear content semantics, yielding excellent results for arbitrary style transfer.