A Hybrid Deep Learning Based Character Identification Model Using CNN, LSTM, And CTC To Recognize Handwritten English Characters And Numerals

Detection and recognition of handwritten English language characters and numerals is facing challenges due to the huge variation and haziness of strokes from one individual to the other, variation in the handwriting style, poor quality of the original document as the paper deteriorate over time, and text in printed documents sit in a straight line whereas humans need not write a line of text in a straight line on white paper. To address these challenges, a deep learning based handwritten text model is proposed that is capable of identifying both characters and numerals from an input image consisting of English language characters and numbers. The proposed model makes use of Convolutional Neural Network (CNN) with Long Short Term Memory (LSTM) method to detect and identify the text. The proposed model has been trained and tested using MNIST dataset consisting of full page images from 3600 writers and 800,000 images. Experimented with different number of epochs, it is found that the proposed model is able to identify the text with an accuracy of about 88.8%.