An End to End Recognition for License Plates Using Convolutional Neural Networks

Abstract

License Plate Recognition technology, as a critical part in Intelligent Transportation Systems, enjoys a broad application. LPR is a technology to read license plates (which include the character and the color) by performing vehicle detection and character recognition on surveillance videos or images. Recently, many different algorithms and models can be applied in license plate detection and character recognition. Yet given that the accuracy of the system will be impacted by image blurring caused by motion blur, low-resolution, low-luminosity, low-contrast and noisy point, concerns have been concentrated in the accuracy uncertainty of complex application scenarios impact. To address these problems, an end-to-end fast solution for number-plate recognition is proposed in this paper. The mainstream algorithm, i.e. Single Shot MultiBox Detector, is employed to auto-detect the location of the license plate. Also, to recognize consecutive characters in the detected license plate area, an end-to-end Convolutional Neural Network classification model is proposed here. The test dataset contains over 6 thousand vehicle images collected in different types of complex scenarios, which include low-resolution, low-luminosity and motion blurring. Besides, the solution proposed in this work is able to achieve 99% accuracy on the test dataset with average computing time being 30 fps, which suggests that the system can meet the requirements of real-time analysis.