Enhancing Handwritten Text Recognition Performance with Encoder Transformer Models

Abstract

Handwritten Text Recognition (HTR) is a critical area in computer vision and natural language processing, aiming to convert handwritten content into machine-readable text. The task poses significant challenges due to the inherent variability in handwriting styles, stroke patterns, character spacing, and writing instruments. Traditional HTR techniques, often based on statistical models or shallow neural networks, frequently struggle to generalize across diverse handwriting samples, leading to suboptimal performance in real-world applications.

This improvement demonstrates a relative gain of approximately 10-15% over traditional RNN and LSTM-based models on the same dataset.