DEVELOPMENT OF A WEB-BASED HANDWRITTEN TEXT RECOGNITION & ARCHIVAL SYSTEM

ABSTRACT

Transcribing old handwritten documents such as notes and invoices to electronic devices has become an essential subject as technology advances to digital databases. Traditional Optical Character Recognition (OCR) systems can only successfully transcribe the printed text. They cannot efficiently deal with handwritten notes due to the disparity in individual cases. This study aims to automate the text recognition process by developing a Handwritten Text Recognition System using Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). This study will account for the disparity of handwriting during analysis and serve as a repository for preserving handwritten data. This study extensively reviews existing systems and research to discover which deep learning architectures are popular in document analysis. The objective was to find the current state of the art in achieving good results. Using PyTorch python library and Google Colab, a Transformer Vision Transformer (Vit) model was used to predict handwritten text images using the IAM dataset for training. This project presents a web-based application using HTML, CSS, Bootstrap, and Flask frameworks. The server side uses SQLite for the database creation and the model embedded into the system. The Transformer architecture could considerably predict image text with a training Character Error Rate of 8.2%. On further evaluation with 364 samples from out-of-dataset images, the system achieved an average accuracy of 3% Character Error Rate. This system can only predict images at the line level and struggles in some cases in the real world, where image quality can be unsuitable for prediction. However, it still serves as a means of transcription and storage for simple documents.