Avaliação de algoritmos baseados em Deep Learning para Localizar placas veiculares brasileiras em ambientes complexos

Abstract

With the increase in the number of private vehicles, we can observe the increase in the number of violations of traffic laws, theft of vehicles and, thus, a better management and traffic control is necessary. A vehicle and its owner are recognized through the unique and required vehicle license plate (LP), and to be inspected and extracted data with greater efficiency, it is recommended to use automated systems for detecting and recognizing vehicle license plates. This work introduce a study and evaluation of algorithms based on Deep Learning to locate Brazilian LPs in complex environments. For the achievement of the experiments, a bank of images of Brazilian LPs was created based on problems like images with different resolution, quality, lighting and perspective of scene. Were used the Deep Learning algorithms YOLOv2 and YOLOv3, which has not yet been studied to the best of our knowledge. In addition, the Tree-structured Parzen Estimator (TPE) algorithm was used to optimize hyperparameters and maximize the performance of selected convolutional neural networks. For the evaluation, the performance metrics were used: prediction time, Intersection over Union (IoU) and confidence rate. The experiments result demonstrate that YOLOv3 presented better performance, obtaining 99.3% of vehicle license plate detection.