Data Augmentation Using Generative Adversarial Network to Improve Classification of Handwritten Words

Date of Award

5-2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computer Science

First Advisor

Proceso L. Fernandez, Jr., PhD

Abstract

Generative Adversarial Networks (GANs) application has resulted in significant advancements in the field of computer vision. This study aims to improve handwriting classification models by developing a data augmentation technique utilizing GAN. The Troika Generative Adversarial Network (T-GAN) for data augmentation was proposed to address the absence of sufficient publicly labeled handwriting datasets. T-GAN is composed of three GANs that work together to generate synthetic images (that appear quite realistic) from three distinct domains by sharing learned weights. T-GAN and other data augmentation techniques were used to augment a set of words from the IAM Handwriting Database with limited instances. The improvements brought by each technique to handwriting classification accuracies in three types of Artificial Neural Networks (ANNs) such as Deep ANN, Convolutional Neural Network (CNN), and Deep CNN were measured and compared. Results showed that the T-GAN-based data augmentation technique outperformed the standard techniques and the techniques that uses another GAN-based model. The paired t-test between the 10-fold cross-validation results validated the improvement brought about by the proposed T-GAN and its superiority as a handwriting data augmentation approach. This clearly shows that T- GAN can be effective at synthesizing good-quality handwriting data by sharing weights and parameters. Finally, the trained CNN and Deep CNN classifiers was improved to 93.62% and 96.36% classification accuracy, respectively, when the generated synthetic data instances from the T-GAN were further enhanced using the pepper noise removal and median filter. Thus, data augmentation using T-GAN, together with the stated two image noise removal approaches could be a promising pre-training combination technique for augmenting handwriting datasets with insufficient data samples.

Link to Full Text

Share

COinS