License Plate Recognition Model For Tilt Correction Based on Convolutional Neural Network

Abstract

The purpose of the study was to discuss how a tilted license plate (LP) affects the accuracy of LP recognition and how to improve a recognition system. The character segmentation on tilted LP usually causes character segmentation to be incomplete or out of range, which leads to a decrease in the accuracy rate of character recognition. We propose a method to improve the accuracy of LP recognition and reduce the prediction model training time for the recognition system. The study has four steps which are LP location, LP correction, character segmentation, and character recognition. Firstly, LP was located and zoomed in with YOLOv4 to reduce irrelevant noise and background value. Secondly, the system analyzed pixel changes of each angle with a horizontal projection and corrected the horizontal tilt angle for the LP. Then, the system used vertical projection to move the upper and lower half pixels of the LP in opposite directions. By analyzing the projection status of each angle, the system then corrected the vertical tilt angle for the LP. Thirdly, the system performed character segmentation on the corrected LP. This was done by extracting each character. Lastly, given more than 9,000 character images from step three, the recognition system with Convolutional Neural Network (CNN) trained the prediction model with the feature selection of the maximum pooling layer. Finally, the recognition system accuracy of predicting the uncorrected LP is 96.1% after 25 epochs, while the recognition accuracy of predicting corrected LP is 99% after 10 epochs. The accuracy of LP recognition was increased from 96.1 to 99% after LP tilt correction. CNN training time was decreased from 25 epochs to 10 epochs.