Content-Aware Urdu Handwriting Generation

Zeeshan Memon, Adnan Ul-Hasan, Faisal Shafait

Abstract: The performance of handwriting recognition systems has undergone significant improvement in recent years. However, the accuracy of these systems for multiple cursive scripts, including Arabic and Urdu, is still limited due to the lack of labeled training data. Handwriting generators are a potential solution to this problem. Previous research on Urdu handwriting generation has primarily focused on generating realistic ligatures using Generative Adversarial Networks (GANs) with common adversarial loss but has not addressed the issue of maintaining content and generated image entanglement. This paper aims to address this gap by proposing a content-controlled training approach for Urdu Handwriting Generation with pre-trained recognizer loss. Our generation model is trained on a diverse set of printed ligatures and then fine-tuned with transfer learning on handwritten images. In this paper, a new metric for evaluating the performance of handwriting generation systems is also suggested, which is specifically tailored to the context of handwriting generation tasks. To our knowledge, this is the first Urdu handwriting generation system that is capable of generating content-controlled images.