Pseudo-coordinates graph convolutional generative adversarial network for art style transfer

The goal of style transfer is to apply the artistic features of a style image to a content image while maintaining the content image's structure. Traditional methods often use CNNs and residual blocks, but their limited receptive field struggles to capture long-range feature dependencies, leading to repetitive local patterns. Residual blocks can also cause interference between style and content representations. To address these issues, we introduce a pseudo-coordinates graph convolutional generative adversarial network (PGC-GAN), which consists of two branches: one for extracting style and another for style transfer. The style extraction branch represents style features as a graph and uses graph pooling to remove redundant information. The style transfer branch encodes these features into pseudo-coordinates, enabling flexible relationships between pixel nodes and long-range feature aggregation without disrupting the content image's structure. Experimental results demonstrate that PGC-GAN significantly improves artistic style transfer compared to existing methods.