A Deep Learning-Based Structural Off-line Approach for the Automatic Recognition of Handwritten Punjabi Characters

Abstract

Aim: This study employs the YOLOv5 deep learning algorithm to crop handwritten Punjabi Gurmukhi letters from scanned photos automatically. Background: The capacity of a computer to read and comprehend handwritten text is known as handwriting recognition technology. The digitization of society and the rising need for automation have led to considerable growth in this technology. Nearly all languages have handwriting recognition software, yet Punjabi is one language without much of it. Objective: By automating character cropping with YOLOv5, the paper's main goal is to reduce the pre-processing load during handwritten character recognition for Punjabi scripts. It also compares the suggested approach with previous work for the same dataset. Method: Using machine learning and deep learning algorithms, some recent research highlights the importance of automatically cropping letters, words, and sentences for handwriting identification. Many languages, except Punjabi, are available for this extensive undertaking. First, boundary boxes were made using a convolutional neural network (CNN) to segment the letters. Additionally, two optimizers—SGD and ADAMW—were used with YOLOv5 to recognize letters. Result: Punjabi alphabets without diacritics reached a maximum accuracy of 98.1% when implemented using YOLOv5 ADAMW, while Punjabi alphabets with diacritics were implemented using an automatic cropping technique and YOLOv5 SGD, yielding 97.4% accuracy. Conclusion: The study emphasized the significance of optimizer selection and how it affects the recognition of Punjabi characters. The paradigm being explained here provides an opportunity for scholars in this discipline.