Text-Based Image Style Transfer and Synthesis
Yifan He, Jian Li, Anna Zhu · 2019
In this work, we propose a novel framework for synthesizing the stylization of text-based images. It is composed of several steps without supervision. Initially, the style image is segmented to foreground and background images. Then, the main color of foreground is accumulated and assigned to the stroke-based binarized geometric shape like text, symbols and icons to be content image. The foreground image is considered as the target style and transferred to the content image by image style transfer neural network. Meanwhile, the blank space in segmented background image is filled by image inpainting-based partial convolution neural network. Finally, the composition of the stylized geometric shape and the complete background image is accomplished by texture synthesis. The features of edge, color, texture, etc., are explored for the stylization of multiple geometric shapes. It explains the rationality of our proposed method on binarized geometric shape style transfer. The extensive experiments on various tasks, such as visual-textual presentation synthesis, icon/symbol rendering and structure-guided image inpainting, demonstrate the effectiveness of the proposed method.