Abstract
Arbitrary artistic style transfer is a research area that combines rationalacademic study with emotive artistic creation. It aims to create a new imagefrom a content image according to a target artistic style, maintaining thecontent's textural structural information while incorporating the artisticcharacteristics of the style image. However, existing style transfer methodsoften significantly damage the texture lines of the content image during thestyle transformation. To address these issues, we propose affinity-enhancedattentional network, which include the content affinity-enhanced attention(CAEA) module, the style affinity-enhanced attention (SAEA) module, and thehybrid attention (HA) module. The CAEA and SAEA modules first use attention toenhance content and style representations, followed by a detail enhanced (DE)module to reinforce detail features. The hybrid attention module adjusts thestyle feature distribution based on the content feature distribution. We alsointroduce the local dissimilarity loss based on affinity attention, whichbetter preserves the affinity with content and style images. Experimentsdemonstrate that our work achieves better results in arbitrary style transferthan other state-of-the-art methods.