CONCEPTUAL MODEL OF AN AI-BASED PATTERN RECOGNITION MANAGEMENT SYSTEM

DOI: 10.31673/2412-4338.2025.038709

Abstract

Recent innovations in computer vision have enabled the introduction of sophisticated recognition models capable of processing complex images and making reliable predictions based on huge amounts of video data. The main problems of training neural networks in the field of computer vision are data bias, which occurs when training computer vision systems on biased sets of information, data interpretability, which complicates the interpretation of decisionmaking processes, and the need for significant computing resources, especially in real-time application scenarios on devices with limited resources. This paper is devoted to the study of methods for improving the efficiency of computer vision, in particular in real time and using minimal computing resources. The paper discusses the principles of convolutional neural networks and compares the characteristics of current recognition algorithms such as R-CNN, RFCN, and YOLO. The main characteristics chosen for comparison are speed, accuracy, and computing resource utilization.

Based on the analysis, the You Only Look Once recognition algorithm was chosen for the presented model of the pattern recognition control system because it combines speed and accuracy, making it suitable for applications that require immediate processing. YOLO performs all calculations simultaneously and allows detecting objects in one pass through the neural network, which significantly increases the speed and significantly differs it from traditional two-stage algorithms. The main stages of the algorithm in the context of the proposed model are defined, including the use of NonMaximum Suppression to discard unnecessary frames and finally determine the optimal frame, and loss functions to optimize accuracy by taking into account coordinates, size, trust, and classification

Keywords: artificial intelligence, pattern recognition, convolutional neural networks, You Only Look Once