A deep learning-based system for massively generating images containing meaningful numerics

We present a deep learning-based system for generating images, such as pictures of electricity meters, in which numbers and letters play an important role. A large amount of image data is often required to build a deep learning-based system for image recognition, so it would be useful to have a system that can automatically generate realistic images. GANs can be used for this purpose, but there are some hurdles to overcome for GANs to create realistic images in which texts are embedded. Most of existing approaches focus on generating either the textual images or the non-textual ones only, not the ones where the textual part is embedded in a small area while still being clearly identifiable. In order to solve this problem, we propose a deep learning-based approach that attempts to learn textual images and non-textual ones independently before generating a set of complete images combined from the learned results. Also, we demonstrate the strengths of the proposed system by providing some empirical results on the electricity meter image data. © The Korean Institute of Electrical Engineers