MA-GAN: A Method Based on Generative Adversarial Network for Calligraphy Morphing

Some applications can be based on the image transfer method in the development of contemporary related fields, which can make the incomplete Chinese font complete. These methods have the ground truth and train on paired calligraphy image from different fonts. We propose a new method called mode averaging generative adversarial network (MA-GAN) in order to generate some variations of a single Chinese character. As the data set of calligraphy font creation through samples of a given single character is usually small, it is not suitable to use conventional generative models. Therefore, we designed a special pyramid generative adversarial network, using the weighted mean loss in the loss function to average the image and the adversarial loss to correct the topology. The pyramid structure allows the generator to control the rendering of the topology at low resolution layers, while supplementing and changing details at high resolution layers without destroying the image topology resulting from the diversity requirements. We compared MA-GAN with other generation models and proved its good performance in the task of creating calligraphy fonts based on the small data sets.