Aprendizado por transferência aplicado na detecção de placas de licenciamento de veículos

Abstract

Vehicle identification is important for various reasons, such as property ownership identification, theft prevention, and even traffic flow control. Various methods are used for vehicle identification, such as numbers engraved on their structures or parts. Among these, the vehicle license plate stands out for its practicality in several functionalities, as it is usually located in easily visible points of the vehicle. However, automatic license plate recognition remains a significant challenge. This is evident when considering the diversity in image capture conditions, including variations in lighting, capture angle, and plate layouts, which can vary between different regions and countries. Many algorithms have been developed to detect and recognize vehicle plates, each with advantages and shortcomings in different situations. In recent years, the scientific community has made advances in methodology and performance, with recent studies showing interesting results in the recognition and segmentation of vehicle license plates using Convolutional Neural Networks (CNNs) in combination with transfer learning techniques. This work addresses the application of transfer learning for the detection and segmentation of vehicle plates using convolutional neural networks. A systematic review was conducted to understand the state of the art in the field. Subsequently, a pre-trained MobileNet architecture with the ImageNet dataset was adapted and trained. Training and validation were performed using images from the UFPR-ALPR dataset, which contains 4500 images of national vehicles. The final model was evaluated using various performance metrics. The results indicated that the combination of transfer learning techniques and fine-tuning significantly improves plate detection accuracy. This study contributes to the advancement of automatic license plate recognition, highlighting the importance of image preprocessing and the use of specialized loss functions for model optimization. [resumo fornecido pelo autor]