A GAN-based Rain Augmentation for Enhancing the Accuracy of the License Plate Detection

Abstract

Nowadays, Automatic License plate recognition (ALPR) solutions are becoming more and more popular and widely applied, from indoor to complex outdoor environments. Despite the many approaches to solving the problem that have been proposed, the ALPR is generally divided into four main phases: image acquisition, license plate detection, segmentation, and character recognition. In that process, the results of the detection and extraction of the license plate region play an important role in influencing the final identification result. The accuracy of this process is greatly influenced by complex environmental conditions, especially rain in tropical countries. In this article, we propose an approach to enhance the accuracy of vehicle license plate detection in rainy conditions based on a solution to augmentation of the training data sets using SyRaGAN. From the Chinese City Parking Dataset (CCPD) training data set, we used SyRaGAN to create five different rain effects corresponding to each original. The original training set and post-enhanced training set will be used to train and evaluate by two state-of-the-art algorithms, You Only Look Once version 5 (YOLOv5) and Faster Region-based Convolutional Neural Network (Faster RCNN). The results demonstrated that models trained from the augmented training dataset gave comparable results to those trained from the original training dataset in traditional test situations but yielded significantly higher efficiency by about 15.95% in YOLOv5 and 10% in Faster RCNN with experimental directions with license plate photos in rainy conditions