Research on License Plate Detection and Recognition Based on Deep Learning

Abstract

With the advent of the era of intelligent transportation, the timeliness and accuracy of license plate(LP) recognition is very important for vehicle management. Traditional LP recognition algorithms rely on fixed scenes and complex image capture systems, no LP recognition algorithm can be widely used in a variety of scenarios; this paper proposes a LP recognition algorithm based on deep learning. First, you only look once(Yolo) v3-tiny with reducing the layers of Yolo v3 is used to roughly locate the LP in the video or image. Then with the landmark detection to precisely detect the LP, and finally recognition the LP with deep convolutional neural network(CNN) end to end. At the same time, in case of the difficulty of data collection, we propose an automatic LP generation algorithm, and we pre-trained base models first, then added real scene data fine-tuning the model for different scenarios to improve the portability and robustness of our models. Through experiments comparison proves that our method has significant advantages in real scenarios with timeliness and high accuracy.