Performance Comparison of various CNN Models for Recognition of Handwritten Telugu Text

Abstract

Deep Learning (DL) stands as a pivotal field in the exploration of pattern recognition, offering unparalleled potential for addressing difficult machine learning challenges. The recognition of characters in the Telugu language using Optical Character Recognition (OCR) presents challenges due to the complex structures of characters, the presence of confusing characters, and overlapping characters.  Convolutional Neural Networks (CNNs) exhibit proficiency in extracting features from training images and determining subtle differences in character shapes. The potency of robust CNN architectures has significantly elevated recognition rates, especially for Indian scripts. Our goal was to evaluate how well CNN based models adapt and perform with accuracy in a challenging script recognition context. In this paper, we introduce a character segmentation algorithm designed to address the challenges posed by overlapping characters, considering the Telugu language-specific features. Our approach involves a preprocessing stage to identify page boundaries and detect words within lines, employing edge detection algorithms. Subsequently, characters are extracted from words on the page using a character segmentation algorithm tailored for the Telugu language and the characters are recognized using trained deep learning models. In addressing the distinctive traits of the training data, we utilize a built based upon the Inception and ResNet models, incorporating adjustments in layers. The model’s performance undergoes thorough validation using a standard dataset, and it is benchmarked against established models in the respective field.