Abstract
In recent years, a variety of systems using deep convolutional neural network (CNN) approaches have achieved good performance on license plate detection and character recognition. However, most of these systems are not stable when the scenes changed, specification of each hierarchical layer to get the final detection result, which can detect multi-scale license plates from an input image. Meanwhile, at the stage of character recognition, data annotation is heavy and time-consuming, giving rise to a large burden on training a better model. We devise an algorithm to generate annotated training data automatically and approximate the data from the real scenes. Our system used for detecting license plate achieves 99.99% mean average precision (mAP) on OpenITS datasets. Character recognition also sees high accuracy, thus verifying the superiority of our method.
