License Plate Character Recognition using Convolutional Neural Network

Abstract

Background: In the last decade, the number of registered vehicles has grown exponentially. With more vehicles on the road, traffic jams, accidents, and violations also increase. A license plate plays a key role in solving such problems because it stores a vehicle’s historical information. Therefore, automated license-plate character recognition is needed. Objective: This study proposes a recognition system that uses convolutional neural network (CNN) architectures to recognize characters from a license plate’s images. We called it a modified LeNet-5 architecture. Methods: We used four different CNN architectures to recognize license plate characters: AlexNet, LeNet-5, modified LeNet-5, and ResNet-50 architectures. We evaluated the performance based on their accuracy and computation time. We compared the deep learning methods with the Freeman chain code (FCC) extraction with support vector machine (SVM). We also evaluated the Otsu and the threshold binarization performances when applied in the FCC extraction method. Results: The ResNet-50 and modified LeNet-5 produces the best accuracy during the training at 0.97. The precision and recall scores of the ResNet-50 are both 0.97, while the modified LeNet-5’s values are 0.98 and 0.96, respectively. The modified LeNet-5 shows a slightly higher precision score but a lower recall score. The modified LeNet-5 shows a slightly lower accuracy during the testing than ResNet-50. Meanwhile, the Otsu binarization’s FCC extraction is better than the threshold binarization. Overall, the FCC extraction technique performs less effectively than CNN. The modified LeNet-5 computes the fastest at 7 mins and 57 secs, while ResNet-50 needs 42 mins and 11 secs. Conclusion: We discovered that CNN is better than the FCC extraction method with SVM. Both ResNet-50 and the modified LeNet-5 perform best during the training, with F measure scoring 0.97. However, ResNet-50 outperforms the modified LeNet-5 during the testing, with F-measure at 0.97 and 1.00, respectively. In addition, the FCC extraction using the Otsu binarization is better than the threshold binarization. Otsu binarization reached 0.91, higher than the static threshold binarization at 127. In addition, Otsu binarization produces a dynamic threshold value depending on the images’ light intensity. Keywords: Convolutional Neural Network, Freeman Chain Code, License Plate Character Recognition, Support Vector Machine