Scene Text Transfer for Cross-Language

Abstract

Scene text transfer for cross-language aims to erase the original scene text and generate another language text image into the original scene text image with the same style, including the style of fonts, colors, size, and background texture. Scene text transfer for cross-language is a challenging problem as the complicated background scene and a huge difference between languages, which demanding high-quality performance for both text transfer and text erasing. In this work, we propose a scene text transfer framework for cross-language which consists of three steps: regional text extraction, style transfer, and scene text combination. The regional text extraction is designed to crop the text region of a natural scene image and transform it to be a rectangle text image. In the second step, a style transfer network is proposed to retain the style of text image and transfer the text content. In the step of the scene text combination, our model combines the rendered text image with the original scene image to produce the final result. In the optimization part, we introduce a novel background consistent loss to improve the performance of background generation. Experiments demonstrate that our framework generates scene text images of higher quality than previous methods.