Reformative Vehicle License Plate Recognition Algorithm Based on Deep Learning

Abstract

As the surge of the number of vehicles in modern cities, it has become more and more difficult to recognize vehicle plates in high accuracy from different shooting situations. Traditional image processing methods like edge or color detection [12, 18] become incapable of detecting plates shooting from side view or with distorted resolution input. Additionally, optical character recognition methods based on graphic features sometimes have high error rate, even when the plate is detected from the front view. Therefore, we present a reformative vehicle license plate recognition algorithm combined with two networks: MobileNet-Single Shot MultiBox Detector (SSD) and Convolutional Neural Network (CNN). We use the SSD to locate the region of interest (ROI) and the CNN to identify the characters on the plate. In the experiments, we show that the SSD and CNN are small and easy to train. What’s more, the results prove that our algorithm performs much higher accuracy than traditional methods in the side view and distorted problems.