A Novel Approach for Text Extraction and Word Segmentation from Handwritten Document Images Using CNN-RNN Technique

Abstract Optical Character Recognition is a technology that takes an optical image of a character as input and generates the corresponding character as output. Its applications span a wide array, encompassing fields such as traffic surveillance, robotics, and the digitization of printed material. Implementation of Optical Character Recognition often involves Convolutional Neural Networks, a widely adopted architecture within the realm of deep learning. Traditional Convolutional Neural Network classifiers excel in learning crucial 2D features within images and subsequently classifying them. This classification process is typically carried out utilizing a SoftMax layer. In this paper, the authors described the optical character recognition by using refined versions of Convolutional Neural Networks and a Recurrent Neural Network classifier. The quality of text recognition was assessed using Character Error Rate and Word Error Rate. Two datasets, IAM and RIMES, were utilized, each divided into training and testing subsets. Accuracy, precision, and recall were calculated based on these divisions. The experimental findings revealed that the Convolutional Neural Networks method achieved notably higher accuracy rates across both datasets, reaching 89.3% and 86%, respectively.