An ALPR System-based Deep Networks for the Detection and

Recognition

Mouad Bensouilah, Mohamed Nadjib Zennir and Mokhtar Taffar

Department of Computer Science, Jijel University, BP 98, Ouled Aissa, 18000, Jijel, Algeria

Keywords:

Computer Vision, Convolutional Neural Networks, Recurrent Neural Networks, GRU, LSTM, Object

Detection and Recognition, OCR, License Plates Recognition.

Abstract:

Automatic license plates reading (ALPR), from images or videos, is a research topic that is still relevant in

the ﬁeld of computer vision. In this article, we propose a new dataset and a robust ALPR system based on the

YOLO object detector of literature. The trained Convolutional Neural Networks (CNN) allow us to extract

features from license plates and label them through Recurrent Neural Networks (RNN) specialized character

recognition. RNN are supported by GRU units instead of LSTM units that are generally used in the literature.

The experiments results were conclusive reaching a recognition rate of 92%.

1 INTRODUCTION

In recent years, deep learning techniques have

achieved good performances in the computer vision

ﬁeld particularly for tasks such as detecting objects

and recognizing their class by offering different deep

network models. These techniques have paved ways

and allowed researchers to use powerful deep learn-

ing models to develop more performant algorithms

and real systems like these used in the ﬁeld of license

plate recognition.

Automatic license plate reading (ALPR) is a com-

mon task these days. It is used for many appli-

cations such as transport, road safety, parking, etc.

A large number of ALPR approaches (Liu et al.,

2011)(Du et al., 2012)(Khan et al., 2017)(Cheang

et al., 2017)(Laroca et al., 2018) have been proposed

in the literature, some of which have been marketed

on real world systems.

Most existing algorithms need segmentation as

a pre-treatment for plate and/or character detection.

Unfortunately, there is no robust segmentation tech-

nique for the variety of constraints to which fonts

are subjected and which cause great variability in

the appearance of characters such as: rotations,

scale changes, blurring, lighting variations and noise

(

Spa

nhel et al., 2017).

In this article we propose a recent approach as

an alternative for the recognition of license plates

without segmentation. It is based on the YOLO

object detector (Redmon et al., 2016)(Redmon and

Farhadi, 2017)(Redmon and Farhadi, 2018) and Re-

current Convolutional Neural Networks (CRNN)(Shi

et al., 2016).

YOLO (Redmon et al., 2016), based on CNN,

is an object detection algorithm, fast, and effective,

and easy to integrate for real applications, with three

versions of YOLO, YOLOv2 (Redmon and Farhadi,

2017) and YOLOv3 (Redmon and Farhadi, 2018).

YOLOv3 is the last version which was optimized

from the previous, with accuracy improved while

maintaining speed and performance. We trained

YOLOv3 with two backbones: Darknet-53 (Red-

mon and Farhadi, 2018) and MobileNets (Andrew

G. Howard, 2017); on the ImageNet pre-training

model. The test results show that the trained YOLOv3

have extremely high recall and precision.

CRNN proposed by (Shi et al., 2016) is a general

framework for character recognition. We applied an

improved CRNN based on CNN for extracting fea-

ture sequences of license plate images, and we used

a 2-layer bidirectional Gated Recurrent Unit (GRU)

(Cho et al., 2014) for labeling sequences. The original

CRNN uses for labeling sequences a popular variant

which is Long Short Term Memory (LSTM) (Hochre-

iter and Schmidhuber, 1997) having more complex

structure than GRU. We used Connectionist Temporal

Classiﬁer (CTC) loss function proposed by (Graves

et al., 2006) during training process. Test results show

that our CRNN have high accuracy than other models

tested on our dataset. It is important to point out that

the Algerian license plates are far more complex than

204

Bensouilah, M., Zennir, M. and Taffar, M.

An ALPR System-based Deep Networks for the Detection and Recognition.

DOI: 10.5220/0010229202040211

In Proceedings of the 10th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2021), pages 204-211

ISBN: 978-989-758-486-2

others, for they have no standard model. Thus, many

kinds of plate patterns are generated from different

fonts, colors, and sizes.

The remainder of the paper is organized as fol-

lows. Section 2 brieﬂy reviews related license plate

recognition literature. Section 3 describes our pro-

posed method. Section 4 introduces publicly our

new ALP dataset, and then we describe in Section

5 the training process on our dataset. The next sec-

tion presents experimental setups and discusses ex-

perimentations and achieved results. Finally, Section

7 concludes the paper.

2 RELATED WORK

In this section, we review recent work that uses deep

learning approaches in the context of ALPR. We dis-

cuss work related to the two phases: plate detection

and plate content recognition.

2.1 Plate Detection

Plate detection is an important phase in the ALPR

system. Numerous research works have proposed

models that speciﬁcally use convolutional neural net-

works (CNN). Li et al. (Li et al., 2018) trained a

CNN capable of detecting the license plate based on

the characters in an image. The detected characters

are grouped in text regions and provided as initial

candidates. Then, false positive plates are removed

by a plate/non-plate CNN classiﬁer. This approach

takes more than two seconds to process a single im-

age when running on a NVIDIA Tesla GPU K40c.

Xie et al. (Xie et al., 2018) proposed a YOLO-

based model called MD-YOLO to detect license

plates rotating angle in addition to coordinates and

its trust value. Before MD-YOLO, Xie et al. ap-

plied a CNN to determine the region of interest in an

image, assuming that some distance inevitably exists

between two license plates. This approach was exe-

cuted in real time on three public datasets. Hsu et al.

(Hsu et al., 2017) proposed custom models of YOLO

and YOLO 2.0 speciﬁcally for the detection of license

plates. These YOLO’s custom models worked better

and were able to process 54 FPS on a ”GeForce GTX

TITAN X” GPU.

2.2 Plate Recognition

After detection, there is an equally important phase

for an ALPR system, which is reading or recognizing

the content of the plate. For this last phase, there are

many papers proposing different approaches. Selmi et

al. (Selmi et al., 2017), for example, proposed a con-

volutional neural network (CNN) with 37 classes. It

contains four convolutional layers, three pooling lay-

ers, one dropout layer and two fully connected layers.

This model is based on a pre-treatment for the extrac-

tion of candidate characters, and then all candidates

are resized to 32 × 32 pixels in gray levels to feed the

model.

On another hand, Li and Shen (Li and Shen, 2016)

took the gray-level image of the detected license plate

and divided it into regions or sub-images of 24 × 24,

then switched all image regions to a convolutional

neural network (CNN) 9-layers with 36 classes. After

the sub-image recognition, the authors eliminated the

false positives with a BRNN with the LSTM units.

Wu et al. (Wu et al., 2018), however, proposed a

DenseNet based model (Huang et al., 2017), with 68

classes for Chinese license plates. DenseNet is a con-

volutional neural network (CNN) with dense blocks,

containing three types of blocks: convoluted block,

dense block and transition block. In their proposed

model, there are a convoluted block, three dense

blocks and two transition blocks to process a gray-

level image of 136 × 136 pixels and recognize the li-

cense plate. Yet, Spanhel et al. (

Spa

nhel et al., 2017)

proposed a convolutional neural network (CNN) with

eight output branches recognizing characters on re-

spective positions with 36 classes. It takes a color

image of 200 × 40 pixels as an input.

3 PROPOSED APPROACH

In ALPR systems ﬁeld, the goal is to build a system

that, from an image, detects license plates and effec-

tively recognizes the existence of writing and charac-

ters in the image. Our approach addresses the prob-

lem of reading license plates in two phases: plate de-

tection followed by plate contents recognition. Figure

1 depicts the overall architecture of our approach.

3.1 Plate Detection

In this ﬁrst phase, we use the YOLO (Redmon

et al., 2016)(Redmon and Farhadi, 2017)(Redmon

and Farhadi, 2018) object detector from literature. It

is a CNN capable of detecting objects in real time and

thus achieving interesting results in terms of compro-

mise (speed/precision) on datasets of the detection of

published objects, such as Pascal VOC (Everingham

et al., 2010) and Microsoft COCO (Lin et al., 2014).

YOLO uses a single CNN that applies to the entire

image. This network divides the image into regions

and predicts the bounding boxs and probabilities for

An ALPR System-based Deep Networks for the Detection and Recognition

205

...

CTC

...

CNN model

19169111316

1- plate detection

Image Input plate detection Cut and Resize Plates

YOLOv3

Grey-level entry plate

with size 128x64

CNN-based sequence feature

extraction

BRNN-based sequence

labeling

CTC-based sequence

decoding

recognized plate

...

[Bounding

Boxs]

[Scores]

[Classes]

2- plate recognition

Figure 1: General scheme of our approach of license plates recognition.

each. These bounding boxs are weighted by predicted

probabilities. YOLOv3 (Redmon and Farhadi, 2018)

is an update to previous versions of YOLO (Redmon

et al., 2016)(Redmon and Farhadi, 2017); proposed

by Redmon and Farhadi in 2018, it is a more advanced

version than the previous one that is more accurate

and faster.

YOLOv3 (Redmon and Farhadi, 2018) predicts

the boxes at 3 different scales. It extracts the features

of these scales using a concept similar to that of pyra-

mid networks. It uses ”k-means clustering” to deter-

mine the history of encompassing areas. The system

uses 9 clusters and 3 arbitrary scales, and then divides

these clusters evenly between scales.

In this paper, we compare the Darknet-53 feature

extractor (backbone) that is proposed for YOLOv3

(Redmon and Farhadi, 2018) to another named Mo-

bileNets (Andrew G. Howard, 2017). MobileNets

was proposed by Howard et al. of Google Inc. in

2017, and it is faster and provides good results. We

assign to the input the images of 416 ×416 pixels and

use the MobileNet-YOLOv3 implementation (Yang,

2019).

3.2 Plate Recognition

For the second phase, we rely on a type of CRNN net-

work proposed by (Shi et al., 2016). CRNN is a hy-

brid neural network model whose architecture is spe-

cially designed to recognize sequence-like objects in

images. It is a combination of a CNN and an RNN.

To extract the sequence of features from an input

image, the CRNN model accommodates a basic con-

voluted neural network (CNN) by keeping convolu-

tion layers and max-polling layers by eliminating the

fully connected layer. The CNN release is the entry

of the recurrent neural networks (RNN) built to make

labeling for each frame of the sequence of features.

3.2.1 CNN for the Extraction of Features

Each image contains features that set it apart from the

rest of the images. To extract these features, we de-

veloped a model inspired by the CRNN model (Shi

et al., 2016) and CNN model it’s the VGG-12. The

VGG-12 (Simonyan and Zisserman, 2014) consists

of 12 layers (convolution and max-polling). It takes

raw image inputs and produces robust feature maps

that contain high-level descriptions of input images.

In other words, it builds representations that bring out

ICPRAM 2021 - 10th International Conference on Pattern Recognition Applications and Methods

206

the properties of objects learned from the images of

inputs. Because a CNN requires input images to be

scaled to a ﬁxed size to meet its ﬁxed input size, we

ﬁx the input images to the size of 128 × 64 pixels.

3.2.2 RNN for Sequence Labelling

A bidirectional recurrent neural network is built above

convolutional layers, such as recurrent layers. RNN

has a strong ability to capture contextual information

in a sequence (Williams and Zipser, 1995). In our

work, we replace the LSTM (Hochreiter and Schmid-

huber, 1997) units with GRU units (Cho et al., 2014)

in both BRNN layers. GRU was introduced in 2014

by Cho et al. (Cho et al., 2014) to allow each recur-

rent unit to adaptably capture dependencies at differ-

ent time scales (Chung et al., 2014). It is similar to

the LSTM unit (e.g., Figure 2), but easier to calculate

and implement (Fu et al., 2016). The GRU has block-

ing units that modulate the ﬂow of information within

the unit, however, without having separate memory

cells (Chung et al., 2014), represented by the follow-

ing equations (1): Initially, for t = 0, the output vector

is h

= 0.

= σ(W

t−1

)

= σ(W

t−1

)

= (1 − z

t−1

+ z

= tanh(W

 h

t−1

))

(1)

Where x

respectively, are the input and output

vectors; z

represent the update and reset gates vec-

tors with U, W two parameter matrices.

(a) Long Short-Term Memory (b) Gated Recurrent Unit

Figure 2: Illustration of (a) LSTM et (b) GRU (Shi et al.,

2016).

In addition to the fact that GRU units perform bet-

ter than LSTM, we have augmented the output vector

classes of GRU from 256 to 512 to improve the recog-

nition rate of our model. Table 1 shows the conﬁgu-

ration of our CRNN model.

3.2.3 Layer Transcription

Transcription is a process of converting the higher

frame predictions made by the coding module into a

label sequence. Mathematically, the transcription pro-

cedure consists of ﬁnding the label sequence with the

highest probability conditioned by predictions about

pre-frames. To accomplish this task, we use the Con-

nectionist Temporal Classiﬁcation (CTC) proposed

by Graves et al. (Graves et al., 2006). The CTC

is based on a procedure inspired by the ”forward-

backward” algorithm, without segmenting the input

sequence before training.

4 TRAINING

We train YOLOv3 with a single class and mini-

batches (16 images per batch) using our dataset (3408

images of different sizes), with the input image sized

to 416 × 416 × 3.We trained the CRNN with eleven

classes (numbers from 0 to 9, plus white special char-

acter). Thus, we trained the model with our dataset

(2179 Algerian license plates) to mini-batches (32 li-

cense plates per batch), where the input image are

sized to 128 × 64 × 1. We used an Adam optimizer

and a learning rate parameter ﬁxed to lr = 0.001 dur-

ing the training step.

5 DATASET

We collected a set of image and video data that

we captured in the municipality of Draria in Alge-

ria using a ﬁxed camera. Then we increased our

dataset by public images from websites that can be

accessed on the Internet (such as ”google images”

”facebook marketplace”

and “Ouedkniss”

). We

have published our dataset on the GitHub reposi-

tory (LPA Dataset, 2019).The vehicle objects of our

dataset exhibit a wide variability of viewpoints and

lighting. Thus, we have built up a learning set of 2408

images, and a test set of 1000 images used to train and

test YOLO detector. Figure 3 presents some samples.

For the license plates recognition, we collect

the plates extracted during the detection step on our

dataset. The latter contains a learning set of 1, 775

plates, and a test set of 404 plates. They were care-

fully and manually annotated with license plate char-

https://www.google.com/imghp

https://www.facebook.com/

https://www.ouedkniss.com/

An ALPR System-based Deep Networks for the Detection and Recognition

207

Table 1: The conﬁguration of our CRNN.

Example Type of layer Parameters

Entry Grey-level image with size 128 × 64

Convolution ﬁlters: 64, kernel: 3 × 3, s: 1, p: 1

Max pooling kernel: 2 × 2, s: 2

Convolution ﬁlters: 128, kernel: 3 × 3, s: 1, p: 1

Max pooling kernel: 2 × 2, s: 2

Convolution ﬁlters: 256, kernel: 3 × 3, s: 1, p: 1

Max pooling kernel: 1 × 2, s: 2

Convolution ﬁlters: 512, kernel: 3 × 3, s: 1, p: 1

Batch normalization

Convolution ﬁlters: 512, kernel: 3 × 3, s: 1, p: 1

Batch normalization

Max pooling kernel: 1 × 2, s: 2

Convolution ﬁlters: 512, kernel: 2 × 2, s: 1, p: 1

BGRU units: 512

Fully Connected units: 11

Softmax classes: 11

acters to allow an accurate assessment of optical char-

acter recognition. Some samples shown in Figure 4.

Figure 3: Image samples used to train the YOLO detector.

Figure 4: Samples of license plates used to train the CRNN.

6 EXPERIMENT AND RESULTS

In this section, we ﬁrst describe a set of measures

commonly used for evaluating deep learning-based

systems. Then, we present our conﬁgurations and the

experimental considerations that we have imposed by

setting the benchmarks and parameters that have al-

lowed us to achieve satisfactory performance and high

rates of detection and recognition of Algerian license

plates.

6.1 Popular Metrics

First we denote by T P, FP, T N, FN respectively

true-positive, false-positive, true-negative and false-

negative.

Precision/Recall. It is a matter of using accuracy or

error rate for the detection of objects. These measures

are summarized in the following equations (2):

Precision =

T P

T P + FP

, Recall =

T P

T P + FN

(2)

IoU: Intersection over Union measures the overlap

between two borders. It is given as equation (3):

IoU =

Area o f Overlap

Area o f Union

(3)

WER/CER. The performance of character recogni-

tion models can be measured by word error rate

(WER) and character error rate (CER). WER is the

report of reading errors calculated at the word level.

ICPRAM 2021 - 10th International Conference on Pattern Recognition Applications and Methods

208

Table 2: Comparison of the performance of YOLOv3 and MobileNet-YOLOv3.

Detection model Average detection speed (sec/Image) Recall Precision AP

YOLOv3 0.050 0.970 0.990 0.969

MobileNet-YOLOv3 0.032 0.960 0.980 0.954

Table 3: Comparison of the performance of CRNN, Wu-DenseNet, and our model.

Recognition Model Input Image Size WAR(%) CER(%)

CRNN (Shi et al., 2016) 128 × 32 × 1 train: 0.06, test: 15.63 train: 0.01, test: 2.46

Wu-DenseNet (Wu et al., 2018) 136 × 36 × 1 train: 0.06, test: 28.29 train: 0.01, test: 4.33

Our Model(BLSTM) 128 × 64 × 1 train: 0.06, test: 9.18 train: 0.01, test: 1.08

Our Model(BGRU) 128 × 64 × 1 train: 0.06, test: 7.94 train: 0.01, test: 0.99

The CER measures the distance of Levenshtein stan-

dardized by the length of the word ground-truth.

6.2 Experiment Setups

We implemented our approach using both Keras and

Tensorﬂow libraries with Adam optimization. The

training and test run using a PC with Intel Core Xeon

CPU and 16 GB GPU (NVIDIA Tesla P100-PCIE).

6.3 Results

Table 2 shows the results obtained by the two mod-

els, tested on our dataset (LPA Dataset, 2019), af-

ter 22,123 training steps. Compared to YOLOv3,

MobileNet-YOLOv3’s detection speed is faster, and it

always meets real-time requirements. While in terms

of Average Precision (AP) performance (Everingham

et al., 2010), YOLOv3 is better than MobileNet-

YOLOv3, which means that the license plates de-

tected by YOLOv3 are closer to the ground truth.

The evaluation criterion for recognition is the ac-

curacy of the license plate, which means that the

recognition is correct when all the characters of a li-

cense plate are correctly recognized. To this purpose

we used two metrics (CER, WAR). Table 3 shows the

results obtained by our model, tested on our dataset

(LPA Dataset, 2019), as well by CRNN (Shi et al.,

2016) and Wu-DenseNet (Wu et al., 2018). All three

models were obtained after 22,188 training steps.

Compared to both models, our model is more efﬁ-

cient for Algerian license plates (CER-0.99%, WAR-

7.94%).

Table 4 shows the results obtained by our model

tested on our dataset (LPA Dataset, 2019) and other

results on the Chinese Dataset-1 that is featured in

Wu-DenseNet (Wu et al., 2018). Chinese Dataset-1

(Wang et al., 2017) contains 203,774 Chinese license

plates for training and 9,986 for testing. Chinese li-

cense plates generally come according to a standard

model and are identical. On the other hand, Algerian

license plates do not have a standard model, causing

the existence of many different models (fonts, colors,

design) of license plates, which has, therefore, accen-

tuates the difﬁculty to recognize the Algerian license

plates (e.g., Figure 5).

Figure 5: Comparisons between Chinese license plates

(top), and Algerian license plates (bottom).

Figure 6 presents some examples of MobileNet-

YOLOv3 detection and CRNN recognition result.

7 CONCLUSION

In this article, we proposed an end-to-end license

plate recognition system based on the YOLO de-

tector and CRNN. For the detection step, we op-

timized some training parameters of YOLOv3 and

MobileNet-YOLOv3, and then we trained a license

plate detection model on datasets with high and uni-

form intra-class variability of plate patterns.

We carried out comparative experiments of

YOLOv3 and MobileNet-YOLOv3, and the results

show that MobileNet-YOLOv3 works better on de-

tection speed while YOLOv3 works better on detec-

tion accuracy. In the recognition part, we designed

and trained our CRNN model which is an improved

RNN. We also carried out comparative experiments

of our CRNN with other models, and the results

show that our system performs better for Algerian li-

cense plates. Experimental results show that the pro-

posed system has achieved top performance in terms

of recognition speed and recognition accuracy, which

An ALPR System-based Deep Networks for the Detection and Recognition

209

Table 4: Comparison of performance on Chinese Dataset and our Algerian Dataset.

Recognition Model Dataset Input Image Size WAR(%) CER(%)

Wu-DenseNet (Wu et al., 2018)

Chinoises (Wang et al., 2017)

Alg

eriennes (LPA Dataset, 2019)

136 × 36 × 1

0.01

28.29

0.001

4.33

Our Model Alg

eriennes (LPA Dataset, 2019) 128 × 64 × 1 7.94 0.99

Figure 6: Samples of MobileNet-YOLOv3 detection and CRNN recognition result.

can fully meet the needs of practical applications. As

perspective, we plan to enrich our dataset by explor-

ing a new ways to improve the recognition of Algerian

license plates.

REFERENCES

Andrew G. Howard, Menglong Zhu, B. C. D. K. W. W. T.

W. M. A. H. A. (2017). Mobilenets: Efﬁcient convolu-

tional neural networks for mobile vision applications.

arXiv.

Cheang, T. K., Chong, Y. S., and Tay, Y. H. (2017).

Segmentation-free vehicle license plate recognition

using convnet-rnn. arXiv preprint arXiv:1701.06439.

Cho, K., Van Merri

enboer, B., Gulcehre, C., Bahdanau, D.,

Bougares, F., Schwenk, H., and Bengio, Y. (2014).

Learning phrase representations using rnn encoder-

decoder for statistical machine translation. arXiv

preprint arXiv:1406.1078.

Chung, J., Gulcehre, C., Cho, K., and Bengio, Y.

(2014). Empirical evaluation of gated recurrent neu-

ral networks on sequence modeling. arXiv preprint

arXiv:1412.3555.

Du, S., Ibrahim, M., Shehata, M., and Badawy, W. (2012).

Automatic license plate recognition (alpr): A state-

of-the-art review. IEEE Transactions on circuits and

systems for video technology, 23(2):311–325.

Everingham, M., Van Gool, L., Williams, C. K., Winn, J.,

and Zisserman, A. (2010). The pascal visual object

classes (voc) challenge. International journal of com-

puter vision, 88(2):303–338.

Fu, R., Zhang, Z., and Li, L. (2016). Using lstm and gru

neural network methods for trafﬁc ﬂow prediction. In

2016 31st Youth Academic Annual Conference of Chi-

ICPRAM 2021 - 10th International Conference on Pattern Recognition Applications and Methods

210

nese Association of Automation (YAC), pages 324–

328. IEEE.

Graves, A., Fern

andez, S., Gomez, F., and Schmidhu-

ber, J. (2006). Connectionist temporal classiﬁcation:

labelling unsegmented sequence data with recurrent

neural networks. In Proceedings of the 23rd interna-

tional conference on Machine learning, pages 369–

376.

Hochreiter, S. and Schmidhuber, J. (1997). Long short-term

memory. Neural computation, 9(8):1735–1780.

Hsu, G.-S., Ambikapathi, A., Chung, S.-L., and Su, C.-P.

(2017). Robust license plate detection in the wild.

In 2017 14th IEEE International Conference on Ad-

vanced Video and Signal Based Surveillance (AVSS),

pages 1–6. IEEE.

Huang, G., Liu, Z., Van Der Maaten, L., and Weinberger,

K. Q. (2017). Densely connected convolutional net-

works. In Proceedings of the IEEE conference on

computer vision and pattern recognition, pages 4700–

4708.

Khan, M. A., Sharif, M., Javed, M. Y., Akram, T., Yas-

min, M., and Saba, T. (2017). License number plate

recognition system using entropy-based features se-

lection approach with svm. IET Image Processing,

12(2):200–209.

Laroca, R., Severo, E., Zanlorensi, L. A., Oliveira, L. S.,

Gonc¸alves, G. R., Schwartz, W. R., and Menotti, D.

(2018). A robust real-time automatic license plate

recognition based on the yolo detector. In 2018

International Joint Conference on Neural Networks

(IJCNN), pages 1–10. IEEE.

Li, H. and Shen, C. (2016). Reading car license plates using

deep convolutional neural networks and lstms. arXiv

preprint arXiv:1601.05610.

Li, H., Wang, P., You, M., and Shen, C. (2018). Reading car

license plates using deep neural networks. Image and

Vision Computing, 72:14–23.

Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P.,

Ramanan, D., Doll

ar, P., and Zitnick, C. L. (2014).

Microsoft coco: Common objects in context. In Euro-

pean conference on computer vision, pages 740–755.

Springer.

Liu, G., Ma, Z., Du, Z., and Wen, C. (2011). The calcula-

tion method of road travel time based on license plate

recognition technology. In Advances in information

technology and education, pages 385–389. Springer.

LPA Dataset, L. P. o. A. D. (2019). License Plates of Alge-

ria Dataset. https://github.com/mouad12345/License

Plates of Algeria Dataset.

Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.

(2016). You only look once: Uniﬁed, real-time object

detection. In Proceedings of the IEEE conference on

computer vision and pattern recognition, pages 779–

788.

Redmon, J. and Farhadi, A. (2017). Yolo9000: better, faster,

stronger. In Proceedings of the IEEE conference on

computer vision and pattern recognition, pages 7263–

7271.

Redmon, J. and Farhadi, A. (2018). Yolov3: An incremental

improvement. arXiv preprint arXiv:1804.02767.

Selmi, Z., Halima, M. B., and Alimi, A. M. (2017). Deep

learning system for automatic license plate detection

and recognition. In 2017 14th IAPR international

conference on document analysis and recognition (IC-

DAR), volume 1, pages 1132–1138. IEEE.

Shi, B., Bai, X., and Yao, C. (2016). An end-to-end train-

able neural network for image-based sequence recog-

nition and its application to scene text recognition.

IEEE transactions on pattern analysis and machine

intelligence, 39(11):2298–2304.

Simonyan, K. and Zisserman, A. (2014). Very deep con-

volutional networks for large-scale image recognition.

arXiv preprint arXiv:1409.1556.

Spa

nhel, J., Sochor, J., Jur

anek, R., Herout, A., Mar

ık,

L., and Zem

ık, P. (2017). Holistic recognition of

low quality license plates by cnn using track anno-

tated data. In 2017 14th IEEE International Confer-

ence on Advanced Video and Signal Based Surveil-

lance (AVSS), pages 1–6. IEEE.

Wang, X., Man, Z., You, M., and Shen, C. (2017). Adver-

sarial generation of training examples: applications

to moving vehicle license plate recognition. arXiv

preprint arXiv:1707.03124.

Williams, R. J. and Zipser, D. (1995). Gradient-based learn-

ing algorithms for recurrent. Backpropagation: The-

ory, architectures, and applications, 433.

Wu, C., Xu, S., Song, G., and Zhang, S. (2018). How many

labeled license plates are needed? In Chinese Con-

ference on Pattern Recognition and Computer Vision

(PRCV), pages 334–346. Springer.

Xie, L., Ahmad, T., Jin, L., Liu, Y., and Zhang, S. (2018).

A new cnn-based method for multi-directional car li-

cense plate detection. IEEE Transactions on Intelli-

gent Transportation Systems, 19(2):507–517.

Yang, A. (2018 (accessed February 29, 2019)). A Keras im-

plementation of Yolov3. https://github.com/Adamdad/

keras-YOLOv3-mobilenet.

An ALPR System-based Deep Networks for the Detection and Recognition

211