Using Synthetic Images for Deep Learning Recognition Process on Automatic License Plate Recognition

The Automatic License Plate Recognition has been the subject of several studies, given its applicability in real world situations (e.g. toll collection, identification of vehicles in parking lots or even for safety issues in vehicle control that cross borders between countries). In this work, we propose an analysis of the influence to retraining a plate recognition model and a deep neural network for object detection, using synthetic plates image databases from the Brazilian licence plates. The proposed data set uses variations of rotation, size and noise to evaluate the robustness. Thus, the influence of the use of synthetic plates images on the accuracy of systems responsible for locating real plates, segmenting the characters and recognizing them was evaluated and in the tests performed there was an increase in accuracy (considering a system trained with real plates) of three stages: character segmentation, letter recognition and number recognition (2.54%, 1.09% and 2, 49% respectively). It stands out the accuracy of 62.47% (in the number recognition step) obtained by a neural network trained exclusively with synthetic data and tested on real plates.