KAGAN: A Chinese poetry style transfer method. (September 2022)

• Record Type: Journal Article
• Authors: Yang, Kai; Zhao, Huihuang; Sun, Yaqi; Liu, Qingyun; Hu, Boxia
• Is Part Of: Computers & electrical engineering. Volume 102(2022)
• Publication Date: 2022-09
• Subjects: Character style transfer -- GAN -- Attention mechanism -- Font generation; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2022.108185
• Languages: English
• ISSNs: 0045-7906

Abstract

Highlights: A novel Chinese poetry style transfer method named KAGAN is proposed. It can better generate single Chinese characters from images of Chinese poetry. We propose an improved character segmentation algorithm based on vertical horizontal projection to segment Chinese characters into single images. We adopt the Smooth L1 loss function and replace the L1 loss to make our method more effective for transferring styles of Chinese characters. The proposed method has a novel, multi-channel discriminator that can generate more accurate images of Chinese characters while reducing the amount of calculation. Abstract: Character style transfer is challenging, especially when working with Chinese characters. Compared with English characters, Chinese characters have a range of structures and font styles and have attracted a lot of attention in recent research. Some GAN-based methods were proposed for Chinese character style transfer; however, these methods were focused on a single character image ignoring Chinese sentences or multiple characters in one image. A Chinese poetry style transfer method is proposed to address the problem based on Chinese character style transfer. The proposed method includes Smooth L1 loss, which is used to generate superior images. A novel key-attention mechanism generative adversarial network (KAGAN) and a multi-channel discriminator are introduced to generate high-quality images of Chinese characters. The experiments demonstrate that our method is better than other transfer methods, and the proposed model has improved nearly 2% from the conventional methods according to the SSIM evaluation metric.