DETERMINATION OF THE DATA VOLUME FOR EFFECTIVE CLASSIFICATION OF VEHICLE LICENSE PLATES

Abstract

This paper presents the results of a study investigating the impact of dataset size and class count on vehicle license plate classification using neural networks. The research was conducted using the YOLOv11 model, a state-of-the-art architecture for object detection tasks. The study comprises four main steps, each designed to assess how different configurations of training data affect the model's performance.

The first step involved training the model on a medium-sized dataset consisting of four classes over 15, 30, and 50 epochs. The second step focused on training the model using the same medium-sized dataset but limited to a single class for 30 epochs. The third step evaluated the model's performance on a smaller dataset over the same set of epochs (15, 30, 50). After each step, detailed analysis of training graphs and performance metrics was conducted to assess how the model's accuracy and loss evolved with different data sizes and epoch counts.

In the final step, a comprehensive comparison of all models was performed, taking into account the overall performance across the various dataset sizes and class configurations. This comparison aimed to identify the optimal training configuration, with a focus on maximizing accuracy and minimizing overfitting, to determine the best approach for vehicle license plate classification.

Through this study, the research highlights the influence of dataset volume and class distribution on model performance, providing insights into how to effectively train neural networks for classification tasks with varying data complexities. The findings underscore the importance of selecting an appropriate dataset size and class count to balance computational efficiency and classification accuracy in real-world applications.