Style Transfer with Subject-Object Separation using Masked Random Adaptive Instance Normalization
Yujia Kang, Changming Zhu · 2023
In order to improve the effect of style transfer and provide a style transfer result that matches human cognition, it is necessary to separate the subject and object of the content image and style image and perform style transfer accordingly. However, in the case of few samples, the effect of style transfer is often unsatisfactory. In order to improve the ability of few-shot style transfer and optimize the visual effect of few-shot style transfer, this paper proposes a Masked Random AdaIN network with subject-object separation. By improving the style transfer method of the AdaIN network and proposing content and style loss computation method for subject-object separation, the style transfer effect of subject-object separation is realized. In order to optimize the visual hierarchy and edge transition of the generated image, this paper also proposes an edge enhancement module. With this module, the generated image can maintain the edge area of the original image and ensure a smooth transition between subject and object, so that the image we get is more in line with human cognition. The experimental results show that images generated by our method have a better subject-object style preservation effect and are more in line with human cognition.