Image Style Transfer for Visual Design Using StarGANv2 and Attention Mechanism

Abstract

In visual design, high-quality facial attribute style transfer and diversified painting style transfer are key to meeting the diverse artistic creation needs. To improve the facial image generation effect in visual design, the study first designed a facial attribute style transfer method based on an improved star shaped generative adversarial network version 2. The generator introduced dense connection modules and random noise modules, and a spectral normalization network was added to the discriminator. On the basis of this model, a painting style transfer model was studied and constructed. Its generator adopted distribution shift convolution and attention mechanism, while its discriminator used spectral normalization and Markov dual discriminator. The findings demonstrated that the attribute generation accuracy of the facial attribute style transfer method was 98.56%, and its average time consumption in practical applications was 71.2ms, significantly better than the comparison model. The PSNR and maximum structural similarity index of the painting style transfer model are 38.68dB and 0.948. Both models had good performance and could provide efficient and high-quality style transfer solutions for visual design. This research holds immense importance in fostering the advancement of artistic creation in the field of visual design.