МЕТОД ЗУПИНКИ ПРОЦЕСУ РОЗПІЗНАВАННЯ РЯДКОВИХ ОБ’ЄКТІВ У ВІДЕО ПОТОЦІ РЕАЛЬНОГО ЧАСУ

Анотація

This article explores the challenge of optimal stopping in real-time string object recognition within video streams, a critical issue for systems with constrained computational resources, such as mobile devices. The problem is framed as a decision-making task that balances the trade-off between computational costs and recognition accuracy. A mathematical model is proposed, introducing a stopping rule based on the expected distance between the current and subsequent integrated recognition results, ensuring efficiency and precision in dynamic environments. The developed algorithm leverages the normalized Levenshtein distance as a metric for assessing recognition accuracy, integrating results using the ROVER algorithm. Experimental validation was conducted on the MIDV-500 dataset, encompassing diverse text areas from identity documents, such as dates, machine-readable zones (MRZ), document numbers, and personal names. Using the Tesseract OCR engine (versions 4.1.1 and 5.5.0), the method demonstrated significant improvements in recognition accuracy and resource efficiency compared to traditional fixed-step and threshold-based approaches. The findings highlight the robustness and versatility of the proposed approach, as it adapts seamlessly to varying text recognition algorithms and computational environments. The article underscores the broader applicability of the method, suggesting extensions to other object types, including images and complex patterns. Furthermore, the integration with machine learning models and dynamic confidence scoring is proposed to enhance decision-making accuracy.