Vehicle License Plate Recognition Based on Class-specific ERs and
SaE-ELM
Chao Gou, Kunfeng Wang, Bo Li, and Fei-yue Wang
Abstract— In this paper, an effective approach to vehicle
license plate recognition based on Extremal Regions (ERs) and
Self-adaptive Evolutionary Extreme Learning Machine (SaE-
ELM) is proposed. In the license plate detection step, some com-
putations including morphological operations, various filters,
different contours and validations are sequentially performed
to extract some image regions as candidate license plates.
Then, accurate character segmentation is achieved through
a proper selection of ERs. In the character recognition step,
the HOG (histogram of oriented gradients) feature vector in
each character region is extracted, and then the characters are
recognized using an offline trained pattern classifier of SaE-
ELM. Experimental results show that our approach works quite
well in complex traffic environments.
I. INTRODUCTION
In recent years, intelligent transportation systems (ITS)
have been widely used to tackle the growing urban problems,
especially traffic congestion and accidents. ITS can roughly
be categorized into intelligent infrastructure systems and
intelligent vehicle systems. With the rapid development of
computer vision technology, more and more vision-based
systems are applied in ITS. For instance, computer vision
systems for vehicle license plate recognition (VLPR) are
used as a core of intelligent infrastructure systems like elec-
tronic payment systems and freeway management systems
[1].
In general, VLPR from traffic images or videos consists
of three processing steps: license plate region detection,
character segmentation, and character recognition.
In order to recognize vehicle license plates, the license
plate regions should first be extracted from a still traffic
image. Accurate detection of license plate regions is essential
to carry on other steps of VLPR. There are two major
methods for locating vehicle license plates: one method is
based on color information [2], [3], while another one is
based on textures or edges of the license plates [4], [5].
In the methods of the first type, many detect the license
plates based on the H component of the HIS color model.
In addition, the color combination of a license plate and its
characters is specific, and this combination occurs almost
This work is supported partly by the National Natural Science Foundation
of China under Grant 61304200, MIIT Project of the Internet of Things
Development Fund: R&D of Real-time Image Recognition Techniques and
Application Systems for Social Security, and the Early Career Development
Award of SKLMCCS under Grant 2F13Q02.
The authors (Kunfeng Wang is the corresponding author) are with
Qingdao Academy of Intelligent Industries, Qingdao 266109, China, and
also with the State Key Laboratory of Management and Control for Complex
Systems, Institute of Automation, Chinese Academy of Sciences, Beijing
100190, China (phone: 86-10-82544791; e-mail: gouchao.cas@gmail.com,
kunfeng.wang@ia.ac.cn).
only in the license plate region [3]. However, the methods
that use color combination to localize license plates may
become invalid when there are regions in the image whose
color information is similar to that of the license plate.
Moreover, license plate detection based on color information
is sensitive to adverse illumination conditions and camera
settings. On the other hand, texture-based methods use high
edge-density areas where color transition occurs dramati-
cally. These methods can detect license plate regions in
relatively simple environments, but can easily be affected
by noises and are computationally complex when there are
many edges in the image.
In the second step, the license plate is segmented to extract
the isolated characters. In [6], the extracted license plate is
rescaled to a template size, while in the template all the
character positions are known. This method is incapable of
dealing with any shift in the extracted license plates. Consid-
ering that the characters and license plate backgrounds have
different colors, some methods [3], [7] project the binary
extracted license plate vertically to determine the starting
and the ending positions of characters, and then project the
extracted characters horizontally to extract each character
alone. The projection method is common and simple, but
is dependent of their accurate positions. It is obvious that
this method needs prior knowledge on the character number
and is sensitive to noises.
Finally, it is difficult to recognize license plate characters
because different camera zooms lead to different character
sizes. Moreover, the extracted characters are very small and
may be similar in their shapes, such as the pairs of S-
5, C-G, and D-0. A large number of character recognition
methods have been proposed, including neural networks
[6], [7], a support vector machine (SVM) in [8], character
templates in [9], [10] and so on. The conventional neural
networks and SVM methods have a high correct rate when
used for character recognition, but they need a long time to
train the model and the recognition procedure is also time-
consuming. The character templates method is unable to deal
with ambiguity of the characters.
In this paper, an effective approach to vehicle license plate
recognition is proposed, based on Extremal Regions (ERs)
and Self-adaptive Evolutionary Extreme Learning Machine
(SaE-ELM) [11].The flowchart of the method is illustrated in
Fig.1. Firstly, top-hat transformation is adopted to preprocess
the input image which helps restrain background noises,
followed by Sobel filter to find the vertical edges and
morphological operations is used to remove blank spaces
between each vertical edge line. Then the coarse LP detection
2014 IEEE 17th International Conference on
Intelligent Transportation Systems (ITSC)
October 8-11, 2014. Qingdao, China
978-1-4799-6078-1/14/$31.00 ©2014 IEEE 2956