An End-to-End Deep Learning Approach for Plate Recognition in Intelligent Transportation Systems

Table 4

Comparison between the proposed methods and the other methods.

Methods	Reported Acc				Plate characters	Number of plates	Image size	Processing time	Method
	Plate detection	Character segmentation	Plate recognition	Overall Acc					D: detection method; R: recognition method
[5]	98.7%	100%	97.6%	96.33%	Persian	10000	Variable	180	D: CCA and RANSAC; R: probabilistic SVM
[10]	99.33%	NR	96.6%	96%	Persian	150		NR	D: color features; R: ANN
[14]	97.3%	NR	94.5%	91.94%	Persian	320		NR	D: edge features (Sobel); R: MLP
[17]	96.93%	98.75%	94.5%	90.45%	Persian	1185		NR	D: morphological operations and Adaboost; R: SAMME
[23]	97.7%	—	98.8%	NR	Chinese	250 K		NR	D: MTCNN; R: MTLPR
[29]	100%	—	96.78%	NR	Chinese	—		—	D: CNN; R: CNN
[35]	NR	99%	NR	NR	Korean	120		NR	D: sliding concentric windows; R: ANN
[30]	100%	NR	98.4%	NR	Chinese	2507		42	YOLOv3 ()
[30]	98.27%	NR	98.1%	NR	Chinese	2507		161	Faster_RCNN_ResNet101 ()
[27]	98.04%	NR	94.12%	NR	Chinese	2049		400	Unified deep neural network
[39]	97.16%	98.34%	97.88%	93.54%	English; Japanese	9026	NR	288	D: improved Bernsen algorithm; R: SVM
[39]	96.5%	NR	89.1%	86%	English	1334	NR	276	D: sliding concentric windows and CCA; R: probabilistic NN
[40]	95.9%	NR	92.3%	90%	Chinese; English	5026		125	D: edge features (Sobel); R: feed forward NN
[41]	97.3%	NR	95.7%	93.1%	English	1176		223	D: salient features; R: self-defined classifier
[42]	97.1%	NR	96.4%	93.6%	English	332		594	D: color features and Hough transform; R: feed forward NN
[43]	94.43%	—	99.37%		Arabic	600		NR	D: DSSN; R: CNN
[44]	97.76%	—	95.05%	NR	Persian	5719	Variable	54.18	D: YOLO-v3; R: YOLO-v3
[45]	99.37%	—	99.53%	98.9%	Brazilian	—		—	D: YOLO-v3; R: CNN
[46]	99.72%	—	87%	—	Jordanian	187200		—	D: YOLO-v3; R: CNN
[47]	98.22%	—	87%	—	English	2049	Variable	—	D: YOLO-v3; R: YOLO-v3
The proposed method	100%	—	99.37%	99.6%	Persian	3 million	Variable	46	D: CNN; R: CNN+LSTM