LSTM Based Character Recognition for Text Extraction in Real-Time Natural Scenes Images

• Published in: 2024 International Conference on Data Science and Network Security (ICDSNS) pp. 1 - 7
• Main Authors: Kant Soni, Vishnu, Chouksey, Praveen, Tandan, S. R., Pimpalkar, Amit, Miri, Rohit, Kumar Nema, Neetesh
• Format: Conference Proceeding
• Language: English
• Published: IEEE 26.07.2024
• Subjects: Accuracy; Adaptation models; Character recognition; Computational modeling; Data mining; Deep Learning; Edge Detection and Intensity Analysis; Image recognition; Long Short-Term Memory; Natural Scenes; Optical character recognition; Real-time systems; Text Detection and Recognition; Text recognition
• DOI: 10.1109/ICDSNS62112.2024.10691042

Extracting text from complex real-world images poses a significant challenge in computer vision due to cluttered backgrounds, diverse fonts, and varying orientations. Traditional methods struggle with accuracy in such scenarios. This research introduces a novel text extraction model, Deep Convolutional Neural Networks (DCNNs), for precise character recognition. By combining Optical Character Recognition (OCR) with Long Short-Term Memory (LSTM) networks, the model effectively handles sequential data and enhances text extraction performance. By integrating OCR with LSTM, the proposed approach aims to overcome the limitations of traditional methods and enhance text extraction performance in challenging scenarios. Evaluation of ICDAR datasets yielded a high average confidence score of 95.50%, highlighting the model's accuracy in challenging text extraction tasks. Real-time testing in Bilaspur, India, demonstrated the model's robustness and adaptability, surpassing human annotations. These results highlight the potential of deep learning, particularly the OCR-LSTM fusion, for accurate text extraction in diverse natural scene images.