Handwritten Recognition of Telugu Characters Using Various Pretrained Convolutional Neural Networks

Telugu handwritten character recognition has seen significant advancements in recent years, particularly in the context of deep learning models such as Convolutional Neural Networks (CNNs). However, recognizing handwritten digits and scripts in multilingual settings, especially in a diverse country like India, presents unique challenges due to the variance in handwriting styles and the complexity of regional scripts. The proposed system addresses the challenges of diverse handwriting styles by leveraging the power of CNNs for automated handwriting recognition. Our first contribution is collecting the Telugu characters dataset. A larger dataset of 33,496 handwritten Telugu characters was collected from schoolchildren and adults of varying ages. We collected a larger corpus of 33,496 locally handwritten Telugu characters to study script variability. The experiments in this paper evaluate six vowels (1,200 images) from the public Kaggle set, serving as a focused benchmark; the larger corpus will be used in follow-up work and released separately. This project presents a deep learning-based approach to recognize handwritten Telugu vowels. To ensure optimal recognition performance, five Convolutional Neural Network (CNN) architectures LeNet, VGG16, ResNet50, DenseNet, and AlexNet were selected and evaluated for the task of Telugu character recognition. These models were rigorously trained, and the Adam optimizer is used to optimize the parameters. These models are evaluated, and their performance is compared. From the analysis, it is found that the VGG16 model showed the best results with a high accuracy of 98.14%.