Image Generation for Printed Character by Representation Learning

With the development of convolutional neural networks, generative models can synthesize really wonderful images. But most of these models are limited in generalization and extensibility. And things become difficult when generating images with multiple specified features. Therefore, this paper introduce an expandable approach to generate images with multiple features. We use our model to generate images including a single character with specified fonts and position, by learning the representations of different features from existing images, and using these representations together. Several structures are proposed to increase the training efficiency and extensibility. Finally, we arrange some experiments and show the performance of our model.