Análise da Acurácia de Detecção de Placa de Veículos com Yolov3 em Imagens com Características de Brilho, Resolução e Desfoque Alteradas

Abstract

Vehicle license plate detection is crucial in various applications such as security, traffic monitoring, and parking management. This study investigates the effect of brightness, resolution, and blur variations on the detection accuracy of the YOLOv3 object detection model. Images from the Universidade Federal de Ouro Preto (UFOP) vehicle database were used, with filters applied to degrade the quality and challenge detection. Average accuracy and error rate graphs were generated to identify points where YOLOv3 model performance declines or fails. It was found that the model's accuracy decreases with a value of 15 as the blur filter kernel, 70% darkening compared to the original image, and 20% of the original resolution (800 pixels x 600 pixels). These findings can aid in developing preprocessing strategies to improve vehicle license plate detection under varying brightness, blur, and resolution conditions.