A New CycleGAN-Based Style Transfer Method

Image style transfer is a research hotspot in computer graphics and computer vision. Especially in recent years, style transfer has been applied in many fields such as art creation, film and television production, and social networking. Therefore, in view of the problem that the image texture generated by the existing style transfer network CycleGAN is blurred, and the style transfer cannot only be performed for a specific foreground, a new CycleGAN-based style transfer method by adding a self-attention layer and semantic segmentation is proposed. The workflow of the network is to first process the image through the optimized CycleGAN to obtain the generated result, send the generated result to the semantic segmentation network to obtain the mask image, perform AND operation based on the mask image to obtain the foreground and background, and finally merge to obtain the output result. The experiments are carried out on two public image datasets, horse2zebra and MNIST. The results show that compared with the original CycleGAN network, this method has a faster texture learning speed and generates clearer textures, and moreover, it solves the problem that the original CycleGAN network cannot perform style transfer only for the foreground.