Automatic Gender Classification from Handwritten Images: A Case Study

Irina Rabaev, Marina Litvak, Sean Asulin, Or Haim Tabibi

Abstract: Using a handwritten sample to automatically classify the writer’s gender is an essential task in a wide range of areas, e.g., psychology, historical documents classification, and forensic analysis. The challenge of gender prediction from offline handwriting can be demonstrated by the relatively low (below 90%) performance of state-of-the-art systems. Despite a high interest within a broad spectrum of research communities, the published works in this area generally concentrate on English and Arabic languages. Most of the existing approaches focus on manual feature selection. In this work, we study an application of deep neural networks for gender classification, where we investigate cross-domain transfer learning with ImageNet pre-training. The study was performed on two datasets, the QUWI dataset, consisting of handwritten documents in English and Arabic, and a new dataset of documents in Hebrew script. We perform extensive experiments, analyze and compare the results obtained with different neural networks. We demonstrate that advanced deep-learning models outperform conventional machine learning approaches that were used in previous studies. We also compare the obtained results against human-level performance and show that the problem is challenging for non-experts.