License plate detection and recognition in unconstrained scenarios

Abstract

As deep learning technology grows by leaps and bounds, license plate detection algorithms have significantly improved their speed and accuracy. Traditional license plate detection generally focuses on the differences in edge, color, or grayscale such simple features. And shows the main problems are lack of vehicle specificity and insufficient robustness. But during these years, algorithms based on deep learning have solved most of the problems and gotten good results. However, because of the changing and complex scenarios in real life, it has to deal with challenges like large-angle license plates, poor lighting environments, and warping license plates. From these unconstrained scenarios, the detection and recognition results are always unsatisfactory.

In this paper, a complete automatic license plate recognition (ALPR) system for unconstrained scenarios is presented, especially focusing on license plates that may be severely warping due to tilt views. A convolutional neural network capable of correcting warping license plates is introduced and finally put into an optical character recognition method to obtain the result. As an additional contribution, separately trained 2 SVMs for Chinese license plates to solve the problem of low accuracy in recognizing Chinese characters. Also, a separate model for the night vehicle data set was trained to improve the accuracy of night vehicle detection.

The effectiveness of this method was proved by testing data sets from three places, between three commercial product with comparing their accuracy and speed.