DEVELOPMENT OF A SYSTEM FOR AUTOMATIC DETECTION AND RECOGNITION OF TEXTUAL SYMBOLS USING NEURAL NETWORKS

Abstract

The aim of the study is to develop software for the automatic detection and recognition of symbols in images using neural networks. The system employs two neural networks: the first analyzes the full-size image, identifies symbols, and determines their coordinates in the QUAD format. The second network processes the extracted image fragments and classifies the symbols based on the training dataset. Text recognition is performed character by character, making the CRAFT model an optimal choice for symbol detection. The symbol recognition model is built using a classical classifier architecture that includes two convolutional layers with pooling and four fully connected layers. The results demonstrate that the combination of the two models ensures a recognition accuracy of 83.87% on the test dataset. Testing confirmed that the main factors influencing accuracy are the quality of the training dataset, learning parameters, and the method of symbol localization. The practical significance of the results lies in the potential application of the developed system for automating text data processing in various fields, such as document management, computer vision, and sign recognition.