Research On Diverse Handwritten Chinese Character Generation Method Based On Generative Adversarial Networ

With the development of deep learning,there have been good advances in Chinese character generation techniques.However,for the problem of generating handwritten Chinese characters,most methods treat them as a special font and use Chinese character generation techniques to generate handwritten characters.One of the main characteristics of handwritten characters is that the same character can have different forms.It is difficult to find two identical handwritten Chinese characters due to factors such as the writer and the writing environment.The method of treating handwritten characters as a font cannot generate personalized and diverse handwritten Chinese characters.In addition,some methods for addressing the lack of diversity in the handwritten Chinese character generation do not use techniques that incorporate structural information of Chinese characters.Instead,they generate handwritten Chinese characters starting from noise and category labels.However,these methods heavily rely on the dataset and cannot generate Chinese characters that do not exist in the dataset.Collecting a complete dataset of handwritten Chinese characters with a diverse range of characters is difficult.Therefore,whether the generated results have diversity and whether they can generate Chinese characters that do not exist in the dataset are the key and difficult points of research in handwritten Chinese character generation.This paper addresses the key and difficult points in the current problem of handwritten Chinese character generation by undertaking the following works:(1)The Handwritten Chinese character GAN(HC-GAN)is designed and implemented,which incorporates the demodulation layer to allow the feature vector to influence the entire image generation process.A special loss function and corresponding training method are employed to encourage diversity in the generated results.The low-dimensional features abstracted from printed Chinese characters are used as the starting point for image generation,enabling the model to generate Chinese characters that are not present in the dataset.After training on a processed dataset,HC-GAN is compared with three other style transfer algorithms that produce diverse results and are extensible,and it is found that HC-GAN can generate handwritten Chinese characters with a certain degree of diversity and accuracy.(2)To address the issue of poor performance in generating complex handwritten Chinese characters by HC-GAN,the generation process was decomposed into three steps: segmentation of the printed Chinese character connected domain,generation of the corresponding handwritten character structure,and concatenation of the handwritten character structure.A Handwritten Chinese character connected domain GAN(HCD-GAN)was designed to address this issue.This model improved the method for calculating image similarity in the loss function by introducing a similarity calculation method based on discrete cosine transform,which reduced the problem of "spotty" generated results.The model was trained step-by-step,and the structure was optimized on this basis.A dataset of connected domains of printed Chinese characters and corresponding structures of handwritten Chinese characters was created for training.Comparison with other methods in terms of accuracy and diversity on this dataset showed that the handwriting Chinese character generation method based on HCD-GAN had achieved certain improvements in accuracy.(3)In response to the problem of decreased diversity in the handwritten Chinese character generation method based on HCD-GAN,we designed and implemented the Handwritten Chinese character Structure Concatenation Network(HCSCN).The HCSCN-based handwritten Chinese character structure concatenation method uses a loop concatenation structure to solve the problem of the uncertain length of handwritten Chinese character structure sequences.After making the corresponding dataset and training the HCSCN,it was found through experiments that the concatenation results of HCSCN were closer to real handwritten Chinese characters.The generated results were evaluated using the evaluation method that integrated the accuracy and diversity of the generated results,and the method based on HCD-GAN and HCSCN obtained good results.