Efficient Model for Handwritten Text Generation with Text Line Awareness

Abstract

We propose a new method for handwritten text generation using generative adversarial networks with multi-class conditional batch normalization, which enables us to use character sequences with variable lengths as conditions for the generator. Compared to existing methods, our method converges faster and its memory requirements are similar regardless of the number of classes, which allows us to train a generative model for a language with a large number of characters, such as Japanese. We also introduce an additional condition that makes the generator aware whether there are characters extending below the baseline or above the mean line in the generated sequence, which helps generate results with well-aligned characters in the text line. Our human evaluation study shows that our proposed method generates handwritten text images that look more realistic and natural.

Publication
The 24nd Meeting on Image Recognition and Understanding (MIRU 2021).