Sunteți pe pagina 1din 4

2017 IEEE Calcutta Conference (CALCON)

Automatic Number Plate Recognition Using CNN


Based Self Synthesized Feature Learning

Madhusree Mondal, Parmita Mondal,


and Nilendu Saha Paramita Chattopadhyay
Department of Electrical and Mechanical Engineering Department of Electrical Engineering
Techno India College of Technology (TICT) Indian Institute of Engineering Science and Technology
Kolkata, India Kolkata, India
{madhusreemondal73028,mondalparmita & paramita_chattopadhyay@yahoo.com
nilendusaha82213}@gmail.com

Abstract - In today's image centric world image processing Unlike the input layer of a conventional NN classifier which is
has become the centre of attention for various real-life fully connected with huge number of weights, CNN employs a
applications. Embarking from car thefts, breaking traffic rules locally responsive input layer where neurons work on small
to contravene restricted expanse, Image Processing has given us a regions of the entire input image and produce local responses,
conviction to put a grinding halt to these malpractices. Inspired
which are further processed in successive steps and finally
by the feature learning capabilities of Convolution Neural
Networks (CNN), the preeminent work is the detection and passed through a fully connected layer with much lower
recognition of the car plate number image which is accomplished dimensions.. Most of the initial applications of CNN are
by dint of Convolution Neural Network (CNN). The number centered on pattern. recognition and classification based on
plate images can be procured by a still camera. Self synthesized image data[19]. CNN has been creating buzz not only for
feature of CNN is capable of recognizing the states of the vehicle object detection but also identifying faces, gesture recognition,
from the number plate with a reasonably high accuracy of 90% Natural language Processing and in Self Driving Cars.
even with very low training size. CNN has proved its robustness Lots of research has been done by the previous researchers in
even with distorted, tilted and illuminated datasets. the field of image processing to detect the car license plate
using gray scale method, localisation, segmentation and
Keywords— Number plate, CNN, Convnet, Deep learning,
recognition[15]. Image processing method using modular
image recognition. system has been reported by Sungkwan Je et. al. [23]. Among
various Deep learning architectures, they have used Caffe net
for recognizing the images. A method for vehicle number
I. INTRODUCTION plate detection and recognition using segmentation and feature
extraction was proposed by Lazrus et al.[4]. Koval et al.[5]
Convolution Neural Networks (CNNs) are categorized as a proposed an algorithm that deblurred the number plate images
special type of feed-forward multi-layer neural network and feed forward neural network technique was embarked for
architecture with deep learning capabilities. Like many of the the recognition process. Ozbay and Ercelebi[6] worked on the
other soft computational techniques, CNN are primarily smearing and dilation technique for automatic vehicle
inspired by the outcomes of a series of rigorous biological identification. Shidore and Narote[7] proposed histogram
experiments to demystify the functionality of visual cortex of equalization followed by dilation and erosion for plate area
mammals. The findings of these experiments changed the extraction. SVM classifiers were used for character
understanding of vision systems in brain and established the recognition. Kumar et al.[8] proposed a method based on edge
presence of clusters of locally connected neurons (receptive detection using Hough Transform. Massoud et al.[9] devised a
fields) which are responsive only to a particular region of the system using dilation, smearing and erosion. Chen and Luo
entire field of vision. Unlike processing on the entire image [10] and Du et al.[2] located license plate using improved
information obtained through the retinal sensors at single prewitt operation. Khalil [11] suggested an approach based on
level, visual cortex utilizes a 3-layer hierarchical approach of moving window with template matching technique. In [16],
locally sensitive simple nerve cells and then ascending in the Vehicle Number Plate Detection for Indian Vehicles, briefly
hierarchy more complex and finally the hyper-complex cells discusses on segmentation and feature extraction by template
each of which works on the activation of the previous layer. matching for recognition process. All the work reported in the
This model is arranged in an echelon formation of feature literature associated with number plate detection are based on
detectors that performs successive processing of the visual hand crafted feature extraction techniques. No one has tried
input pertaining the dominant view of visual processing in the self generated feature learning techniques for information
visual cortex and inspired the computational model of CNN. extraction from vehicle number plate.

978-1-5386-3745-6/17/$31.00 ©2017 IEEE 378


2017 IEEE Calcutta Conference (CALCON)

Keeping the above viewpoints in mind the proposed work dataset was generated where each class contained 200 images.
aims to exploit the feature learning capability of convolution Number plates with distortion, tilt, and illumination at
neural network (CNN) for extracting information, more different angles have been considered here for giving it a
particularly, the state or province associated with the realistic effect. Some of the sample training sets are presented
registration of the owned vehicle from the number plate. in Fig 3.
II. PROPOSED METHODOLOGY
CNN involves four key operational layers like i) Convolution,
ii) ReLu (Non-linearity), ii) Maximum pooling and iv) Fully
Connected layer. The schematic representation of the CNN
based proposed methodology is shown in Fig. 1. The process
of Convolution includes the formation of the featured layer by
multiplying the values in the filter with the original pixel
values of the image. Inputs from the convolution layer are
smoothened to reduce the sensitive nature of the filters
towards noise and variations. The process of pooling mainly Fig. 3. Typical Training Samples of Number Plates .
reduces the size of the image, or the color contrast across red,
green, blue (RGB) channels. CNN is compatible to a wide The entire algorithm has been implemented using MATLAB
variety of complex activation functions to model signal R2017a with GPU specification 940MX.
propagation. One of the common function is the Rectified
Linear Unit (ReLU), is favourable for its faster training speed. III. RESULTS AND DISCUSSION
The last layers in the network are fully connected, where the Fig. 4. The performances of the proposed algorithm has been evaluated with
neurons of preceding layers are connected to every neuron in devised training sets of various sample size. The classification accuracies
subsequent layers. with different training samples are depicted in Fig 4. ,WLVFOHDUIURPWKHILJXUH
 WKDW WKH &11 KDV OHDUQLQJ DELOLW\ IURP YHU\ ORZ WUDLQLQJ VDPSOHV &ODVV
ZLVH DFFXUDFLHV KDYH EHHQ SUHVHQWHG LQ WKH WDEOH , )LJ  DQG  GHSLFW WKH
LPSDFW RI LPDJH VL]H DQG NHUQHO VL]H RQ FODVVLILFDWLRQ DFFXUDFLHV 7KH GDWD
GLVWULEXWLRQRQWKHIHDWXUHspace are furnished in fig. 7.

Fig. 1. Schematic representation of CNN Based algorithm.

Fig 4: Variation in accuracy with increasing training data.

TABLE I.
CLASSES KERNEL SIZE
Training Size ACCURACY
& IMAGE SIZE
KERNEL SIZE: TRAINING
TAMIL NADU 94.9
Fig. 2. Flow Chart of the CNN Based algorithm . 5X5 30%
JHARKHAND 99.4

The flow chart of the CNN based algorithm is shown in Fig 2. MAHARASHTRA
IMAGE SIZE: TESTING
99.4
For simplicity we have considered only four classes namely 28X28 70%
West Bengal, Maharashtra, Tamil Nadu, and Jharkhand. To WEST BENGAL 96.6
check the robustness of the algorithm a large number of

379
2017 IEEE Calcutta Conference (CALCON)

IV. CONCLUSION
In this paper, an attempt has been made to recognize states
(province) of a vehicle from the number plate using self-
generated features on CNN. The initial results obtained from
this study are very encouraging, even with a very small amount
of training samples. Though the present investigation shows a
promising future of CNN for this kind of automation
applications, but actual challenge lies on real time
implementation of CNN.
REFERENCES

[1] C. T. Hsieh, C. Liang-Chun, K. M. Hung, and H. Hsieh-Chang, "A real-


time mobile vehicle license plate detection and recognition for vehicle
monitoring and management," in Joint Conferences on Pervasive
Computing (JCPC), pp. 197-202,2009.
[2] Y. Du, W. Shi and C. Liu, "Research on an Efficient Method of License
Plate Location" Physics Procedia, vol. 24, , pp. 1990-1995., 2012
[3] S. Kumar, S. Agarwal and K. Saurabh, "License Plate Recognition
System for Indian Vehicles", International Journal of Information
Fig. 5. Variation in accuracy with increasing kernel size. Technology and Knowledge Management, vol. 1, no. 2 , pp. 311-
325,2008
[4] A. Lazrus, S. Choubey and G.R. Sinha, "An Efficient Method of Vehicle
Number Plate Detection and Recognition", International Journal of
Machine Intelligence, vol.3, no. 3, pp. 134-137,2011.
[5] V. Koval, V. Turchenko, V. Kochan, A. Sachenko, and G.
Markowsky,"Smart License Plate Recognition System Based on Image
Processing Using Neural Network" in IEEE International Workshop on
Intelligent Data Acquisition and Advanced Computing Systems, Lviv,
Ukraine, , pp. 123-12,. Sep.2003
[6] S. Ozbay and E. Ercelebi, "Automatic Vehicle Identification by Plate
Recognition", in World Academy of Science, Engineering and
Technology, pp. 222-225, 2005
[7] M. M. Shidore and S. P. Narote, "Number Plate Recognition for Indian
Vehicles", International Journal of Computer Science and Network
Security, vol. 11, no. 2, pp. 143-146, 2011.
[8] P. M. Kumar, P. Kumaresan and Dr. S. A. K. Jilani, "The Real Time
Vechicle License Plate Identification System", International Journal of
Engineering Research and Development, vol. 2, no. 4, pp. 35-39, 2012.
[9] M.A. Massoud, M. Sabee, M. Gergais and R. Bakhit, "Automated New
License Plate Recognition in Egypt" Alexandria Engineering Journal,
vol.52, no.3, pp. 319-326,2013.
[10] R. Chen and Y. Luo "An Improved License Plate Location Method Based
Fig. 6. Variation in accuracy with increasing image size. on Edge" in 2012 International Conference on Applied Physics and
Industrial Engineering, 2012.
[11] M.I. Khalil, "Car Plate Recognition using the Template Matching
Method", International Journal of Computer Theory and Engineering, vol.
2, no. 5, pp.1793-8201,2010.
[12] P. Anishiya and Prof. S. M. Joans, "Number Plate Recognition for Indian
Cars using Morphological Dilation and Erosion with the Aid of Ocrs" in
International Conference on Information and Network Technology,
Singapore, 2011.
[13] S. Lawrence, C. L. Giles, A. C. Tsoi and A. D. Back, "Face Recognition:
A Convolutional Neural-network Approach", IEEE Transactions on
Neural Networks, vol. 8, no. 1, , pp. 98 ± 113,1997.
[14] Y. le Cun, "Generalization and network design strategies", Technical
Report CRG-TR-89-4 June 1989
[15] A. Amirkhani Shahraki, A. Ebrahimi Ghahnavieh and Seyyed Abdollah
0LUPDKGDYL ³/LFHQVH 3ODWH ([WUDFWLRQ IURP 6WLOO ,PDJHV´ WK
International Conference on Intelligent Systems, Modelling and
Simulation, 2013.
[16] +DQLW.$NVKD\*³9HKLFOH1XPEHU3ODWH'HWHFWLRQ6\VWHPIRU,QGLDQ
9HKLFOHV´ ,((( ,QWHUQDWLRQDO &RQIHUHQFH RQ &RPSXWDWLRQDO
Fig. 7. Features generated for the different classes. Intelligence & Communication Technology
.

380
2017 IEEE Calcutta Conference (CALCON)

[17] CVPR 2014 TUTORIAL ON DEEP LEARNING FOR VISION.[online]. [21] Jianfeng Xu, Zhibin Chen, DQG 6KDRID /L ³&RORU DQDO\VLV IRU &KLQHVH
https://sites.google.com/site/deeplearningcvpr2014/ FDUSODWHUHFRJQLWLRQ´5,663SS-1316
[18] W. Naiguo, Z. Xiang wei, Z. -LDQ ³/LFHQVH 3ODWH 6HJPHQWDWLRQ DQG [22] <DQSLQJ %DL =KHQ -LQ  ³3UHGLFWLRQ RI 6$56 (SLGHPLF E\ %3
5HFRJQLWLRQ RI &KLQHVH 9HKLFOH %DVHG RQ %311´ th International QHXUDO QHWZRUNV ZLWK 2QOLQH SUHGLFWLRQ VWUDWHJ\ &KDRV´ 6ROLWRQV 
Conference on Computational Intelligence and Security 2016 Fractals, Vol 26/2 pp 559-569 , 2005
[19] 3UDEKDNDU3$QXSDPD35HVKPL65³$XWRPDWLF9HKLFOH1XPEHU Plate [23] -H +RQJ+DQK 1JX\HQ DQG -XQK\XN /HH ³,0$*( 5(&2*1,7,21
'HWHFWLRQ DQG 5HFRJQLWLRQ´ ,QWHUQDWLRQDO &RQIHUHQFH RQ &RQWURO 0(7+2' 86,1* 02'8/$5 6<67(06´ ,QWHUQDWLRQDO &RQIHUHQFH
Instrumentation, Communication and Computational Technologies on Computational Science and Computational Intelligence, 2015
(ICCICCT), 2014 [24] .6LPRQ\DQDQG$=LVVHUPDQ³9HU\GHHSFRQYROXWLRQDOQHWZRUNVIRU
[20] Abbas M. Al-Ghaili, Syamsiah Mashohor, Abdul Rahman Ramli, and large-VFDOH LPDJH UHFRJQLWLRQ´ &R55 YRO DEV 
$O\DQL ,VPDLO ³9HUWLFDO-Edge-Based Car-License-Plate Detection [Online]. Available: http://arxiv.org/abs/1409.1556
0HWKRG´,(((7UDQVDFWLRQVRQ9HKLFXODU7HFKQRORJ\YROQR
2013

381

S-ar putea să vă placă și