High Cursive Complex Character Recognition using GAN External Classifier

2025-09-03

AI Summary

Recognition accuracy of handwritten characters deteriorates significantly under high cursive connectivity and severe shape deformation. Method: This paper proposes ADA-GAN, a framework integrating Generative Adversarial Networks (GANs) with an external classifier; it selects high-fidelity synthetic samples based on discriminator confidence and incorporates adversarial noise for controllable data augmentation. Contribution/Results: ADA-GAN overcomes the limited generalization capability of conventional CNNs on complex handwriting patterns by augmenting training data with semantically consistent synthetic samples. It substantially enhances model robustness against highly cursive and structurally intricate characters. Experiments demonstrate that, under identical conditions, ADA-GAN achieves a 12.7% absolute accuracy improvement over the baseline CNN on a high-complexity character subset, with reduced performance variance—validating its effectiveness and stability.

Technology Category

• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Machine Learning: Adversarial Learning & Robustness
• Natural Language Processing: Generation

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Abstract

Handwritten characters can be trickier to classify due to their complex and cursive nature compared to simple and non-cursive characters. We present an external classifier along with a Generative Adversarial Network that can classify highly cursive and complex characters. The generator network produces fake handwritten character images, which are then used to augment the training data after adding adversarially perturbed noise and achieving a confidence score above a threshold with the discriminator network. The results show that the accuracy of convolutional neural networks decreases as character complexity increases, but our proposed model, ADA-GAN, remains more robust and effective for both cursive and complex characters.

Problem

• Classifying highly cursive complex handwritten characters
• Addressing accuracy decline in CNNs with character complexity
• Enhancing robustness using GAN-based data augmentation

Innovation

• GAN with external classifier for cursive characters
• Adversarially perturbed noise for data augmentation
• ADA-GAN model maintains robustness with complexity