Handwriting Recognition Using Wasserstein Metric in Adversarial Learning

Abstract

Deep intelligence provides a great way to deal with understanding the complex handwriting of the user. Handwriting is challenging due to its irregular shapes, which vary from one user to another. In recent advancements in artificial intelligence, deep learning has unprecedented potential to recognize the user’s handwritten characters or words more accurately than traditional algorithms. It works well on the concept of the neural network, many algorithms such as convolutional neural network (CNN), recurrent neural network (RNN), and long short term memory (LSTM) are the best approaches to get high accuracy in handwritten recognition. However, much of the existing literature work lacks the feature space in a scalable manner. A model consisting of CNN, RNN, and transcription layer called CRNN and the Adversarial Feature Deformation Module (AFDM) is used for the affine transformation to overcome the limitation of existing literature. Finally, we propose an adversarial architecture comprised of two separate networks; one is seven layers of CNN with spatial transformation networks (STN), which act as a generator network. Another is Bi-LSTM, with the transcription layer working as a discriminator network and applying Wasserstein’s function to verify the model effectiveness using IAM word and IndBAN dataset. The performance of the proposed model was evaluated through different baseline models. Finally, It reduced the overall word error and character error rate using our proposed approach. © 2022, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd.