Advancing Optical Character Recognition for Low-Resource Scripts: A Siamese Meta-Learning Approach With PSN Framework
Open Access
- 2 December 2024
- journal article
- research article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 189651-189666
Abstract
With the increasing demand for digitization, Optical Character Recognition (OCR) systems play a vital role in digitizing physical manuscripts. Several methods have been successfully deployed in the OCR domain. However, they often face challenges when dealing with low-resource regional scripts because of the limited training data and complex structure of characters. In such a scenario, Siamese Network (SN) meta-learning offers a promising solution for this problem by enabling quick adaptation to new tasks with minimal training data. Despite the success of SNs in various classification tasks, the traditional SN architecture seeks a compelling upgrade to improve its ability to distinguish between similar-looking characters of regional scripts. In this research paper, we propose a novel Priority-Smart Network (PSN) framework for traditional SN architectures, which can easily be incorporated into existing CNN backbone and improve their ability to identify characters in low-resource regional scripts. Furthermore, we propose the Enhanced Differential Edge Detection (EDED) preprocessing strategy explicitly designed for OCR tasks. With rigorous investigation and evaluation of three benchmark low-resource script datasets, we establish the effectiveness of our proposed techniques. Our experimentation results showcase significant advancements in character recognition accuracy and robustness, emphasizing the potential of SN combined with the PSN framework and EDED strategy for improving OCR systems in low-resource script.Keywords
Funding Information
- DST-INSPIRE Fellowship (Innovation in Science Pursuit for Inspired Research), funded by the Department of Science and Technology (DST), Government of India
This publication has 40 references indexed in Scilit:
- Explaining Siamese networks in few-shot learningMachine Learning, 2024
- Deep Siamese Network with Handcrafted Feature Extraction for Hyperspectral Image ClassificationMultimedia Tools and Applications, 2023
- ICCL: Independent and Correlative Correspondence Learning for few-shot image classificationKnowledge-Based Systems, 2023
- DMHC: Device-free multi-modal handwritten character recognition system with acoustic signalKnowledge-Based Systems, 2023
- Person-Specific Face Spoofing Detection Based on a Siamese NetworkPattern Recognition, 2022
- BDNet: Bengali Handwritten Numeral Digit Recognition based on Densely connected Convolutional Neural NetworksJournal of King Saud University Computer and Information Sciences, 2022
- Few-Shot Learning for Tamil Handwritten Character Recognition Using Deep Siamese Convolutional Neural Network2021
- A comprehensive survey on Indian regional language processingSN Applied Sciences, 2020
- Odia character recognition system: A study on feature extraction and classification techniquesMaterials Today: Proceedings, 2020
- The Omniglot challenge: a 3-year progress reportCurrent Opinion in Behavioral Sciences, 2019