Segmentation-free Bangladeshi License Plate Recognition ...

Segmentation-free Bangladeshi License Plate Recognition Using YOLO withHeuristic Bounding Box Refinement

Mahmudul Hasan Bhuiyan Muhammad Anwarul Azim Mohammad Khairul [email protected] [email protected] [email protected]

Farah Jahanfarah [email protected]

Department of Computer Science and EngineeringUniversity of Chittagong, Chittagongon-4331, Bangladesh

ABSTRACT

Automatic License Plate Recognition is the core ofIntelligent Transportation System because of its di-verse set of applications. There are some onlineAPIs such as OpenALPR, Sighthound, etc. avail-able for recognizing license plates of some coun-tries; however, no API is available for recogniz-ing Bangladeshi license plates. The aforemen-tioned APIs cannot even localize Bangladeshi li-cense plates. In this paper, we employ the FastYOLO detector for the first time to Bangladeshi li-cense plate recognition. We also propose an aspectratio-based bounding box refinement technique thatexperimentally outperforms the conventional wayof fixed padding. Various performance evaluationmetrics show that our proposed system not onlyprovides impressive performance but also outper-forms the state-of-the-art methods of license platerecognition. Additionally, we also introduce anannotated dataset containing samples of differentchallenging situations which will help to reduce thescarcity of benchmark datasets in this domain.

KEYWORDS

Automatic License Plate Recognition, YOLO,Transfer Learning, Bounding Box Refinement.

1 INTRODUCTION

Automatic License Plate Recognition (ALPR)works as the heart of Intelligent Transporta-tion System [1]. It has a lot of applications in-cluding automatic parking management, auto-matic toll collection, traffic rules violation de-tection and law enforcement, access and bor-der control and so on [2]. A typical ALPR

system consists of three major steps: (i) Li-cense plate localization; (ii) Character segmen-tation, and (iii) Character recognition. The li-cense plate localization step is responsible forlocalizing the License Plate (LP) in a givenimage, character segmentation isolates individ-ual characters from the detected LP and finally,character recognition aims at recognizing iso-lated characters. Because of the importanceof successful localization of the LP, many ap-proaches first detect the vehicle from an in-put image and then the LP of the vehicle toreduce false positive responses [3]. Despiteits enormous real-life applications, ALPR isstill one of the challenging problems in com-puter vision [4]. The difficulties of the prob-lem lie in the highly complicated background,the variability of LPs, and random photograph-ing conditions such as illumination, distortion,occlusion or blurring [5]. Due to these chal-lenges, most of the works are done under someconstraints such as specific distant cameras orviewing angles, good lighting conditions, sim-ple backgrounds and considering certain re-gions or types of vehicles [3].According to the statistics of Bangladesh RoadTransport Authority (BRTA), a governmentagency, in March 2018 the number of regis-tered vehicles in this country is over 3.4M. It iseasy to realize that ALPR system is a must tomaintain the traffic and ensuring accountabilityof traffic rules violation in such a highly pop-ulated country. To ease traffic management,the Bangladesh government has standardizedthe LP by making it compulsory to use Retro-Reflexive license plates for the ease of visu-alization in different light conditions. There

1

International Journal of New Computer Architectures and their Applications (IJNCAA) 11(1): 1-9The Society of Digital Information and Wireless Communications, 2021 ISSN 2220-9085 (Online); ISSN 2412-3587 (Print)

are some variations in the color and size of theLPs. A typical LP is shown in Figure 1.

Figure 1. Different parts of a typical Retro-Reflexivelicense plate used in Bangladesh.

The first row of a LP contains one of the 64District Codes (DC) which specifies the dis-trict where the vehicle is originally registered,Metropolitan Code (MC) specifies whether thedistrict is a metropolitan or not, and VehicleClass Character (VCC) specifies type of the ve-hicle. There are 31 types of vehicle defined bythe BRTA; for example ‘ja’ for public minibus,’ga’ for motor car of 1301 to 2000 Cubic Cen-timeters (CC) engine etc. If a vehicle is not reg-istered in any metropolitan city, then the MCremains absent. A total of six digits in the sec-ond row specify the registration number of thevehicle.

2 RELATED WORKS

Bangladeshi vehicle images collected fromcertain distances and angles were consideredin [6]. They used empirical techniques to lo-cate the region of interests, horizontal and ver-tical projections for segmentation and neuralnetwork for recognition.In [7], the authors applied Clip Limited Adap-tive Histogram Enhancement (CLAHE) to pro-cess the input images of Bangladeshi LP beforelocalizing. Character segmentation was doneby using some character features such as pixelconnectivity, line space and size etc. Severaldistinctive and mathematical features from thecharacters were used for character recognition.For isolating the character in LPs, a special fea-ture of Bengali called “Matra” (horizontal lineat the upper portion of a character) was usedin [8]. Morphological operations were applied

for extracting the LP and correcting the skewand rotation. The “Matra” was used to segmentindividual characters and finally, a neural net-work was used for recognition.The conventional three stage pipeline of ALPRwas employed in [9]. Morphological opera-tions such as dilation and erosion were used forextracting candidate regions. The regions werebinarized with adaptive thresholding and thenConnected Component Analysis (CCA) wasemployed to segment the characters. Finally,Support Vector Machine (SVM) was used forclassification.In [10], an ALPR system that required to spec-ify the shape and dimension of the LP was pro-posed. The authors compared different contourdetection and object extraction methods. Thepossibility of using content based CCA for LPlocalization is discussed in [11].LP localization in various hazardous condi-tions was considered in [2]. Different tech-niques such as noise removal, contrast en-hancement, tilt correction, rain removal filteretc. were used to correctly localize the licenseplates.In [4], the authors fine-tuned three deep Con-volutional Neural Network (CNN) for local-ization, segmentation and recognition. Theynamed the API as ‘Sighthound’ and showedthat it outperformed several benchmark ALPRsystems when evaluated on two benchmarkdatasets of USA and European License plates.The paper in [3] proposed an ALPR systembased on the state-of-the-art You Only LookOnce (YOLO) object detection model. Theirmodel performed better than ‘OpenALPR’ APIand ‘Sighthound’ on SSIG dataset which is thelargest public dataset of Brazilian LPs. Theyalso introduced a more robust and richer pub-lic dataset for the ALPR task named ‘UFPR-ALPR dataset’.Chandra et al. [12] used CLAHE, Otsu’sthresholding and Contour Analysis forBangladeshi LP localization and projectionfor character segmentation. Google TesseractOptical Character Recognition (OCR) Enginewas used to recognize the MC and two otherCNNs were trained to recognize the VCC and

2


digits.Morphological operations and CNN basedALPR system for recognizing BangladeshiLPs was proposed by Rabbani et al. [13]. Sizeand aspect ratio of the LPs were used for LPlocalization and CCA is used for character seg-mentation. Finally, CNN was used for charac-ter recognition.In [14], the authors presented a system to de-tect the Bangladeshi LP with white backgroundonly. They applied morphological analysis andshape verification using distance to border vec-tors (DtBs) for LP detection, and CNN for rec-ognizing the segmented characters.The authors of [1] employed YOLOv3 andResNet-20 architecture for recognizing onlyDhaka metropolitan vehicle LPs. They usedYOLO for detecting LP and its digits andResNet-20 architecture for the VCC. As onlyDhaka metropolitan city vehicles were consid-ered, they omit the part of DC and MC.A BLPR system on video stream was presentedin [15]. MobileNetV2 based CNN is used forvehicle detection from video frames. ThenYOLOv4 object detection model is used for de-tecting the LP and Google Tesseract OCR en-gine is used for recognizing the characters. Thedataset was created by capturing video of thetraffic in Dhaka city.A three step BLPR system was proposed in[16]. First, the LP is detected using YOLOv3,then a segmentation method based on BreadthFirst Search (BFS) and Flood-fill algorithmwas applied and finally a CNN was used forcharacter recognition. In their system, theyonly considered the LP of Dhaka metropolitancity.A two-step deep CNN based approach is pre-sented in [17]. They fine-tuned Faster R-CNNand SSD MobileNet for both the LP detectionand Character recognition. They presented theperformance of different combinations of theselected CNNs, however the size of the testdata was very small.

3 METHODOLOGY

The classical methods have been used inBangladeshi License Plate Recognition

(BLPR) are restrictive and not robust. Basedon our investigation, no such work was foundon the whole BLPR pipeline using pre-trainedCNN detectors. No benchmark dataset hasbeen referenced in the literature for BLPRand hence researchers used their own datasets.In our work, we use the YOLO detector forthe whole BLPR pipeline pre-trained on thePASCAL VOC dataset. We also propose adataset 1 for the BLPR so that future workscan show comparative performance on thisdataset.

3.1 System Design

We propose a Detect Plate and Detect Charac-ters design for the BLPR pipeline. This impliesthat we use one CNN for LP detection and an-other for character detection. This design pro-vides the best compromise among processingspeed, complexity, and accuracy. The segmen-tation is done directly by CNN as a subprocessof character detection. The absence of the seg-mentation step reduces the error propagationand results in a less overall system error. TheFast YOLOv2 416× 416 version is chosen forboth the LP detection and character detection.It is fast enough for processing real-time videoand provides high accuracy [18]. Both mod-els can be fine-tuned separately to improve theperformance of the corresponding step.The proposed BLPR pipeline consists of twoFast YOLOv2 models named ‘Plate DetectionCNN’ and ‘Character Detection CNN’. Theplate detection CNN detects the LP from agiven vehicle image which has been passed tothe character detection CNN for detecting thecharacters. It detects all the contents of anLP, i.e., the DC, MC, VCC, and the registra-tion number. The architecture of our system isshown in Figure 2.

3.2 Dataset Preparation

The experimental dataset has been formed bycombining the dataset of [12] and the data col-lected by ourselves. It contains a total of 500images with samples of Dhaka, Chattogram

1http://bit.ly/BLPR-Dataset

3


Figure 2. Architecture of the proposed system.

and Sylhet divisions. We split the dataset into atraining set of 400 and a test set of 100 images,every image contains a vehicle image with LP.The test set has been formed in such a way thatit contains samples with different challengessuch as slanted, blur, low light, shadow andnight images, etc. and it contains number ofsamples of each of the cities in accordance withtheir participating portion in the training set.Some samples from the dataset are shown inFigure 3.YOLOv2 supports annotation in PASCALVOC format where an XML annotation file isgenerated for an image to represent the con-tents of the image by means of some fixedtags. All the images of the dataset are man-ually annotated. In the annotation of vehicleimages, only one bounding box represents thewhole license plate and in the annotation ofclosely cropped license plate images. Each ofthe bounding boxes represents any of the DC,MC, VCC or digit.

3.3 Training

As the number of data samples in our trainingset is not so high, fine-tuning the whole net-work is not feasible. Instead, we have mod-ified the last two layers of CNNs and fine-tuned them with our data. Vehicle images havebeen used for training the plate detection CNNwhere the annotations give the bounding boxes

(a)

(b)

Figure 3. Few examples from the dataset. (a) 12 imagesfrom the training set and (b) 12 images from the test set.

of the LPs and the character detection CNNhas been trained on closely cropped licenseplate images where bounding boxes representthe contents of the license plates. The Table 1shows the modified architecture of the CNNs.

Table 1. Network architecture of the CNNs of ourproposed approach.

Plate Detection CNN Character Detection CNNLayer Layer Type Size/Stride Filters Output Filters Output

0 Input 416× 416× 3 416× 416× 31 Convolutional 3× 3 16 416× 416× 16 16 416× 416× 162 Maxpool 2× 2/2 208× 208× 16 208× 208× 163 Convolutional 3× 3 32 208× 208× 32 32 208× 208× 324 Maxpool 2× 2/2 104× 104× 32 104× 104× 325 Convolutional 3× 3 64 104× 104× 64 64 104× 104× 646 Maxpool 2× 2/2 52× 52× 64 52× 52× 647 Convolutional 3× 3 128 52× 52× 128 128 52× 52× 1288 Maxpool 2× 2/2 26× 26× 128 26× 26× 1289 Convolutional 3× 3 256 26× 26× 256 256 26× 26× 25610 Maxpool 2× 2/2 13× 13× 256 13× 13× 25611 Convolutional 3× 3 512 13× 13× 512 512 13× 13× 51212 Maxpool 2× 2/1 13× 13× 512 13× 13× 51213 Convolutional 3× 3 1024 13× 13× 1024 1024 13× 13× 102414 Convolutional 3× 3 1024 13× 13× 1024 1024 13× 13× 102415 Convolutional 1× 1 30 13× 13× 30 165 13× 13× 16516 Softmax

To facilitate the training and prediction pro-cess, each of the possible Bengali objectspresent in a LP has been mapped to their cor-responding English counterparts. The Table 2shows the mapping of Bengali objects to theirrespective English counterparts.

4


Table 2. Mapping of Bengali objects to their Englishcounterparts.

The training of the plate detection network isstopped after 50 epochs with an error of 0.96to avoid over-fitting. Whereas, the charac-ter detection network has been trained for 100epochs and the loss remain as high as 5.19.The reason behind the high loss of the networkmight be the high number of character classesand the rareness of most of them in the trainingsamples. The learning curve of the networksare shown in Figure 4.

Figure 4. Learning curves of the CNNs.

3.4 Choice of Confidence Thresholds

The confidence threshold value has been usedto control the false positive and false negativerate in object detection. The prime concern ofthe plate detection network is to detect all thelicense plates as there is no backup system todetect the undetected plates. The only way tofind an undetected license plate is to manuallyexamining the result of detection which is con-trary to the aim of the system. If the thresh-old value is set low and a false license plate isdetected, it does not affect the system perfor-mance significantly because the character de-tection network can’t detect any character from

that. Contrarily, for the character detection net-work, we aim to detect only the characters onwhich the network is fairly confident. If thecharacter detection network makes a false pos-itive prediction, then there is no way to dis-cover the error without manually checking allthe predictions which is also contradictory tothe aim of the system. After testing with dif-ferent confidence threshold values, the confi-dence threshold value for plate detection CNNhas been set to 0.2 and the value for characterdetection CNN has been set to 0.25.

3.5 Bounding Box Refinement

Our LP detector attains an average Intersectionover Union (IoU) of 0.7298, however, it stillcutdowns some characters. The aspect ratioof an object remains the same as the distanceand the size of an image changes unless thereis a change in viewpoint. It has a close rela-tionship with the IoU of the predicted bound-ing box. Our experiments have shown that theaspect ratio of bounding boxes with high IoUranges roughly from 1.7 to 2.3. If the aspectratio is less than 1.7, in most cases width ofthe bounding box is smaller in comparison tothe height and if the aspect ratio is greater than2.3, the width of the bounding box is larger incomparison to the height. Therefore, by tak-ing these values as thresholds, we have foundthe best combination of margin for each of thecases as shown in Figure 5. This aspect ratiobased bounding box refinement improves lo-calization accuracy in place of adding a fixedmargin.

4 EXPERIMENTAL RESULT

We have performed our experiment using anIntel Core i5 7th generation 2.5GHz computerwith 8GB RAM. Our system has been imple-mented in Python 3 on a Windows platform.The required APIs are TensorFlow, OpenCV3.0 and Darknet which is an open-source neu-ral network framework.

5


Figure 5. Flow chart of bounding box refinement.

4.1 Performance of Bounding Box Refine-ment

Application of our bounding box refinementhas increased the average IoU by 0.0254 i.e.,from 0.7298 to 0.7552. Whereas, the maxi-mum average IoU obtained is 0.7396 with afixed margin of 20 pixels. This increased IoUcontributes in obtaining better recognition ac-curacy by reducing cutout objects as pointedout in Table 3.

Table 3. Performance improvement of license platedetection with bounding box refinement.

Metric Before AfterAverage IoU 0.7298 0.7552No. of cutout objects 17 11

4.2 Performance of Plate Detection CNN

The performance of plate detection CNN hasbeen evaluated on the test set of the vehicleimages. The mean Average Precision (mAP)gives the average precision for different objectclasses. The obtained mAP for plate detectionis 94.95% and it is the same as the average pre-cision in this case as we have only one class.Figure 6 shows some True Positive (TP) andFalse Negative (FN) detection results.

4.3 Performance of Character Detection

The performance of the character detectionCNN has been evaluated on the test set of

(a)

(b)

Figure 6. Results of plate detection in some challengingsamples. (a) FN detections and (b) TP detections.

closely cropped license plates. The Figure 7shows the number of True Positive (TP) andFalse Positive (FP) for different object classes.

Figure 7. Number of TP and FP samples for each objectclass.

The mAP along with average precision valuesfor different object classes are shown in Fig-ure 8. The mAP of our character detection net-work is 64.21%. Since the mAP is the arith-metic mean of the average precision for differ-ent classes, it is adversely affected by the low-est precision obtained for the rare classes.Our character detection network cannot rec-ognize some vehicle class characters whichare very rare (< 15 samples) in the trainingdataset. The mAP is greatly affected by these

6


Figure 8. The mAP of character detection CNN.

classes. If only the classes detected by the net-work are considered for calculating the mAPvalue, then the mAP value will be as high as98.12% which is much higher than the overallmAP value of 64.21%. Figure 9 shows correctof characters by our method in some challeng-ing samples.

Figure 9. Examples of character detection for somechallenging samples.

4.4 Performance Comparison

Table 4 compares the detection performance ofour system and the state-of-the-art method pro-posed in [12].

Table 4. Comparison of license plate detection accuracyof our proposed method and Chandra et al. [12].

System Correctly Detected Accuracy (%)Chandra et al. [12] 64 64Our System 95 95

The authors of [12] used Google’s TesseractOCR for recognizing the DC and MC. Ourcharacter detection CNN itself can do the same

job. However, we only compare the perfor-mance of VCC and digits recognition of bothsystems in Table 5.

Table 5. Comparison of character and digit recognitionaccuracy (in %) of our proposed method and [12].

System Character Recognition Digit RecognitionTotal Success Accuracy Total Success Accuracy

Chandra et al. [12] 64 50 78.13 384 328 85.42Our System 95 80 84.21 570 556 97.55

Run time is an important factor in developingreal time ALPR systems. The comparison ofthe running time of our approach and that pre-sented in [12] is given in the Table 6.

Table 6. Comparison of running time of our proposedmethod and [12].

System Total time (s) Time per image (ms)Chandra et al. [12] 21.62 216.2Our System 54.43 544.3

The deep CNNs are specially developed forrunning on GPUs. Different experiments showthat running on GPU can speed up CNN pro-cessing by a factor of 10 or more based on thetype of GPU used [19]. Our system is expectedto provide real time performance when run ona GPU like the original YOLO. However, ourcurrent system is slower than the system usedin [12] when run on CPU.

5 CONCLUSION AND FUTURE WORK

This paper presents a Fast YOLO basedBangladeshi License Plate Recognition systemwith an aspect ratio based bounding box re-finement technique to improve the detection ofthe original YOLO network. Two pretrainedCNNs are incorporated for license plate local-ization and recognition where the last two lay-ers of them are fine-tuned for transfer learn-ing the general features. Experimental resultsshow that the proposed approach outperformsthe system proposed in [12] in terms of ac-curacy and also demonstrates the capabilityof meeting the robustness requirement of thereal-world operating environments. A PAS-CAL VOC format annotated dataset is also pre-sented to address the scarcity of benchmark

7


datasets in this domain. Two separate networksare trained and fine-tuned using vehicle im-ages and carefully cropped license plate im-ages for license plate localization and licenseplate recognition respectively.We aim to improve our research findings byexploiting the latest YOLO models and otherCNN models such as variants of R-CNN whichare well suited for directly reading text from animage. Other possibilities of improving bound-ing box refinement using license plate featuressuch as edge and border can also be investi-gated. The challenge posed by a high varietyof rare vehicle class characters and scarcity ofreal image can be addressed by carefully fabri-cating artificial license plate images.

REFERENCES

[1] S. Abdullah, M. Mahedi Hasan, andS. Muhammad Saiful Islam. Yolo-basedthree-stage network for bangla licenseplate recognition in dhaka metropolitancity. In 2018 International Conference onBangla Speech and Language Processing(ICBSLP), pages 1–6, 2018.

[2] Samiul Azam and Md Monirul Islam.Automatic license plate detection in haz-ardous condition. Journal of Visual Com-munication and Image Representation,36:172–186, 2016.

[3] Rayson Laroca, Evair Severo, Luiz AZanlorensi, Luiz S Oliveira, Gabriel Re-sende Goncalves, William RobsonSchwartz, and David Menotti. A robustreal-time automatic license plate recog-nition based on the yolo detector. In2018 International Joint Conference onNeural Networks (IJCNN), pages 1–10.IEEE, 2018.

[4] Syed Zain Masood, Guang Shu, AfshinDehghan, and Enrique G Ortiz. Licenseplate detection and recognition usingdeeply learned convolutional neural net-works. arXiv preprint arXiv:1703.07330,2017.

[5] H Li, P Wang, and C Shen. Towards end-to-end car license plates detection andrecognition with deep neural networks.corr abs/1709.08828 (2017).

[6] M Shadab Mashuk, M Arif Majid,Naimul Basher, and Tarif Riyad Rah-man. Automatic detection of banglacharacters in bangladeshi car registra-tion plates. In 2010 Second Inter-national Conference on ComputationalIntelligence, Modelling and Simulation,pages 166–171. IEEE, 2010.

[7] Nahian Alam Siddique, Asif Iqbal, FahimMahmud, and Md Saifur Rahman. Devel-opment of an automatic vehicle licenseplate detection and recognition systemfor bangladesh. In 2012 InternationalConference on Informatics, Electronics& Vision (ICIEV), pages 688–693. IEEE,2012.

[8] Raiyan Abdul Baten, Zunaid Omair,and Urmita Sikder. Bangla licenseplate reader for metropolitan cities ofbangladesh using template matching. In8th International Conference on Electri-cal and Computer Engineering, pages776–779. IEEE, 2014.

[9] Md Azher Uddin, Joolekha Bibi Joolee,and Shayhan Ameen Chowdhury.Bangladeshi vehicle digital licenseplate recognition for metropolitan citiesusing support vector machine. InProc. International Conference on Ad-vanced Information and CommunicationTechnology, 2016.

[10] Charbel Fares. Intelligent license platerecognition. Inter. J. New Comp. Archi-tectures Applic, 3:54–69, 2013.

[11] Ismail A El Sayad, Mohammad ABazzoun, Hawraa I Younes, Laila MGhoteime, and Samih Abdulnabi. Con-tent based image retrieval using multi-modal data based on cca. InternationalJournal of New Computer Architectures

8


and Their Applications, 8(3):119–125,2018.

[12] Sudipto Chandra, Md Al-amin Nowshad,M Jahirul Islam, et al. An automatedsystem to detect and recognize vehiclelicense plates of bangladesh. In 201720th International Conference of Com-puter and Information Technology (IC-CIT), pages 1–6. IEEE, 2017.

[13] Golam Rabbani, Mohammad AminulIslam, Muhammad Anwarul Azim,Mohammad Khairul Islam, andMd Mostafizur Rahman. Bangladeshilicense plate detection and recognitionwith morphological operation and con-volution neural network. In 2018 21stInternational Conference of Computerand Information Technology (ICCIT),pages 1–5. IEEE, 2018.

[14] Prashengit Dhar, Sunanda Guha, TonoyBiswas, and Md Zainal Abedin. A sys-tem design for license plate recognitionby using edge detection and convolu-tion neural network. In 2018 Interna-tional Conference on Computer, Commu-nication, Chemical, Material and Elec-tronic Engineering (IC4ME2), pages 1–4.IEEE, 2018.

[15] Md Onim, Saif Hassan, Muhaiminul Is-lam Akash, Mahmudul Haque, andRaiyan Ibne Hafiz. Traffic surveillanceusing vehicle license plate detection andrecognition in bangladesh. arXiv preprintarXiv:2012.02218, 2020.

[16] Md Mesbah Sarif, Tanmoy Sarkar Pias,Tanjina Helaly, Md Sohel Rana Tutul,and Md Nymur Rahman. Deep learning-based bangladeshi license plate recogni-tion system. In 2020 4th InternationalSymposium on Multidisciplinary Studiesand Innovative Technologies (ISMSIT),pages 1–6. IEEE, 2020.

[17] Tariqul Islam and Risul Islam Rasel.Real-time bangla license plate recogni-tion system using faster r-cnn and ssd: A

deep learning application. In 2019 IEEEInternational Conference on Robotics,Automation, Artificial-intelligence andInternet-of-Things (RAAICON), pages108–111. IEEE.

[18] Joseph Redmon and Ali Farhadi.Yolo9000: better, faster, stronger. InProceedings of the IEEE conference oncomputer vision and pattern recognition,pages 7263–7271, 2017.

[19] Saide Isilay Baykal, Deniz Bulut, andOzgur Koray Sahingoz. Comparing deeplearning performance on bigdata by usingcpus and gpus. In 2018 Electric Electron-ics, Computer Science, Biomedical En-gineerings’ Meeting (EBBT), pages 1–6.IEEE, 2018.

9


Segmentation-free Bangladeshi License Plate Recognition ...

Documents

Transcript of Segmentation-free Bangladeshi License Plate Recognition ...