CLF-Net: A Few-Shot Cross-Language Font Generation Method

Qianqian Jin, Fazhi He, Wei Tang

Abstract: Designing a font library takes a lot of time and effort. Few-shot font generation aims to generate a new font library by referring to only a few character samples. Accordingly, it significantly reduces labor costs and has attracted many researchers’ interest in recent years. Existing works mostly focus on font generation in the same language and lack the capability to support cross-language font generation due to the abstraction of style and language differences. However, in the context of internationalization, the cross-language font generation task is necessary. Therefore, this paper presents a novel few-shot cross-language font generation network called CLF-Net. We specifically design a Multi-scale External Attention Module (MEAM) to address the issue that previous works simply consider the intra-image connections within a single reference image, and ignore the potential inter-image correlations between all reference images, thus failing to fully exploit the style information in another language. The MEAM models the inter-image relationships and enables the model to learn essential style features at different scales of characters from another language. Furthermore, to solve the problem that previous approaches usually generate characters with missing or duplicated strokes and blurry stroke edges, we define an Edge Loss to constrain the model to focus more on the edges of characters and make the outlines of generated results clearer. Experimental results show that our CLF-Net is outstanding for cross-language font generation and generates better images than the state-of-the-art methods.