Convolutional Neural Networks and YOLOv5 for the Detection of License Plates in Digital Photographic Images

In this research, a dataset of two thousand images was obtained, which were taken through a speed camera to vehicles in the city of Lima, Peru, with the purpose of detecting license plates. Therefore, a distribution of this dataset was made: 70% for training (1400 images), 20% for validation (400 images), and 10% for testing (200 images). To visualize and analyze the behavior and convergence of the loss function, a distribution was made using the cross-validation method with the following three folds: 1400, 840, and 280, with two optimization methods (stochastic gradient descent with momentum and Adam), both optimization techniques were used for training from scratch and transfer learning. The results obtained demonstrated that the loss function converges better with the first optimization method. We used precision metric, too.