An Efficient CNN-based Approach for Automatic Brazilian License Plate Recognition

Automatic License Plate Recognition (ALPR) is increasingly becoming the target of many studies in computer vision due to its great applicability in urban environments. In Brazil, a new type of license plate called Mercosul started being used in 2020, and only few works proposed to include the detection of this kind of plate. To address this issue, we trained models using approximately sixty-five thousand unlabeled images from the Federal Highway Police (PRF) captured by the institution's radars. Using three Convolutional Neural Networks (CNN) it was possible to build an ALPR system with an accuracy of 97.91% in the UFPR-ALPR Dataset, 95% in the SSIG SegPlate Dataset and 79.77% in the UFRN Dataset, presented in this paper. These results outperform the most reliable studies for Brazilian license plates so far.