APPLICATION OF NEURAL NETWORKS FOR AUTOMATED TEXT AND SYMBOL RECOGNITION IN CYBERSECURITY TASKS

Abstract

This work presents the development and investigation of an Optical Character Recognition (OCR) system for low-quality text using machine learning methods. To address the task, two grayscale image datasets were created: the first consisting of isolated English alphabet characters and digits (4,960 images of 250×50 pixels), and the second containing fragments of meaningful text from the book The Hunger Games (4,010 images of 680×50 pixels). To enhance model robustness, the images were distorted using blur and digital noise functions. The OCR model was built based on a combination of Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) with a Connectionist Temporal Classification (CTC) layer for sequence correction. Training was conducted over 70–80 epochs with a 9:1 split for training and validation datasets. A comparative analysis was carried out between the developed system and Tesseract OCR. Experimental results demonstrated that the proposed model achieves superior recognition performance on low-quality images, particularly those affected by digital noise, whereas Tesseract OCR significantly loses accuracy under such conditions. The results confirm the effectiveness of hybrid neural network architectures for recognizing distorted text. Future work will focus on recognizing multi-line meaningful text and improving the model's resilience to various types of visual distortions.