Style transfer network for complex multi-stroke text

Neural style transfer has achieved success in many tasks. It is also introduced to text transfer, which uses a image generate transferred images with textures and shapes consistent the semantic content of reference image. However, when structure complex, existing methods will encounter problems such as stroke adhesion unclear edges. This affect aesthetics generated bring lot extra workload designers. paper proposes an improved network for complex multi-stroke texts. We use shape-matching GAN baseline perform following modifications: (1) morphological methods, erosion dilation, are processing; (2) SN-Resblock added network, BCEWithLogits loss texture network; (3) AdaBelief optimizer adopted constrain structure. Further, new dataset traditional Chinese characters constructed train model. Experimental results show that proposed method outperforms state-of-the-art on both simple characters. shown our increases readability text.