Handwritten text recognition using deep learning techniques: A survey

Abstract

HTR (Handwritten Text Recognition) is the automated process of converting handwritten text into digital text, holding immense value in digitizing historical records and facilitating data entry. Through a combination of image processing and HTR systems decode handwritten characters and words. Pre-processing techniques increases image quality by reducing noise and correcting orientation, while models, like “convolutional neural networks” and “recurrent neural networks”, extract features and capture sequence patterns. Effective HTR models demand diverse training datasets and involve supervised learning to align predicted text with actual transcriptions. Post processing tools, including language models and spell-checkers, refine recognition outcomes. HTR's significance spans historical archive digitization, automated form processing, and aiding individuals with disabilities. Challenges persist in deciphering complex handwriting and handling degraded documents. The integration of deep learning advances HTR, enhancing its accuracy and efficiency, thereby expanding access to handwritten texts and enabling their digital search ability and edit ability. The outcome of this endeavor is a robust and user-friendly tool capable of converting handwritten notes, letters, manuscripts, and other textual materials into editable digital text. This project contributes significantly to bridging the gap between analog and digital information, offering immense potential for archival preservation, data accessibility improved productivity across domains.