The style transfer model of illustration images based on multi-scale CycleGAN

Traditional image style transfer methods cannot preserve the content and structural features of the original image while maintaining a specific style. To preserve the semantic information of the original image, a multi-scale cycle-consistency generative adversarial network model is developed. This model can enable innovative style transformations while maintaining the original artistic characteristics of illustrations. This model can better capture and integrate detailed features of different styles by performing style transfer at different resolution levels. The results showed that the proposed model improved the inception score by 1.755 and 0.122 respectively compared to the other two methods, indicating a significant improvement in image generation quality and superiority in image generation. When the low-level texture feature loss, adversarial loss, and high-level concept feature loss were removed, the Frechet inception distance value significantly increased from 73.72 to 102.28, an increase of approximately 38.74%, emphasising the role of these components in the model. The model proposed in this study achieves diverse style transfer and can maintain high image quality when generating stylised images, providing artists and designers with greater creative inspiration and choice space.