| Handwritten Chinese characters generation has been an important research topic in the field of computer vision since the digital age.Traditional methods mainly focus on the local features of the characters,nevertheless lack the knowledge of the global data distribution.Generative Adversarial Networks(GANs)are able to automatically learning the distribution of the real data,and generate samples indistinguishable from the real ones.It is important that the generated characters have both similar handwritten style and accurate strokes.Due to the complexity of structures and styles in handwritten Chinese characters,however,the results generated by previous GANs are often in wrong strokes and hard to be recognized.In this paper,we propose two new GANs models to solve the difficulties in generating handwritten Chinese characters:(1)This paper proposes the Optimized Feature Matching Conditional GAN(OFMcGAN)model to generate handwritten Chinese characters based on our optimized feature matching algorithm and structure consistency loss.The original feature matching algorithm learns more distribution information of the real data by matching the L2 norm differences of features between the real and fake data;however,the generated results are often blurry and lack of diversity.To solve the problems,we propose the OFM algorithm,which uses L1 norm rather than L2 norm.Both the Generator and Discriminator minimize the OFM loss during the training process.This paper proposes structure consistency loss to further reduce the structure differences between the real and generated characters.Experimental results show that,compared to the original feature matching algorithm,OFM-cGAN effectively improves the quality of the generated characters.OFM-cGAN is superior to other GANs on benchmark datasets as well.(2)This paper proposes OFM-CycleGAN model to transform standard fonts into handwritten Chinese fonts using few existing samples.The original CycleGAN is able to generate Chinese characters with handwritten style,however,due to the complexity of handwritten Chinese characters,the generated results contain a lot stroke errors,thus cannot matching the corresponding labels.To further improve the generation quality,we propose OFM-CycleGAN based on the OFM algorithm.This paper employs OFM algorithm in both forward and backward mapping processes in CycleGAN to provide more data information for the model.Compared to traditional methods,OFMCycleGAN includes no stroke extraction or other complex techniques.Experimental results show that,compared with the original CycleGAN,OFM-CycleGAN effectively improves the quality of the generated characters,and successfully generates different handwritten fonts from different standard fonts.Various experimental results show that,with our proposed methods,the generated results are effectively improved with better generation quality and diversified writing styles.We exceed others on the accuracy of strokes and structures as well. |
