License Plate Detection Using Neural Networks
Luis Carrera
1
,MarcoMora
1
,Jos´eGonzalez
2
, and Francisco Aravena
2
1
LesFousduPixel
Image Processing Research Group
Department of Computer Science
Catholic University of Maule
Casilla 617, Talca, Chile
lcarrera@lfdp-iprg.net, mora@spock.ucm.cl
http://ganimides.ucm.cl/mmora/
2
TUTELKAN
Casilla 33, Talca, Chile
{joseluis.gonzalez,francisco.aravena}@tutelkan.com
http://www.tutelkan.com
Abstract. This work presents a new method for license plate detec-
tion using neural networks in gray scale images. The method proposes
a multiple classification strategy based on a Multilayer Perceptron. It
consists of many classifications of one image using several shifted win-
dow grids. If a pixel belongs or not to the licence plate is determined
by the most frequent answer given by the different classifications. The
result becomes more precise by means of morphological operations and
heuristic rules related to shape and size of the license plate zone. The
whole method detects the license plates precisely with a low error rate
under non-controlled environments.
1 Introduction
The license plate recognition (LPR) is a complex matter widely written about.
The problem itself is how to recognize license plate characters of a front or
rear image of a vehicle. In general, the LPR system has the following parts: the
acquisition of the image, the image preprocessing, the detection of the license
plate, the segmentation and the characters recognition [1].
The focus of this paper is the detection step, in other words, determining the
zone where the license plate is. In the literature many techniques for this step
have been reported. A segmentation method based on thresholds is proposed in
[2]. Usage of Fuzzy Logic is shown in [3]. Edge detection by means of gradient
and morphological techniques are presented in [4]. The image scanning using
adaptive windows, considering heuristics of statistical descriptors is proposed in
[1]. The horizontal and vertical projection is presented in [5]. The line detection
using the Hough transformation is proposed in [6]. Learning techniques and
Neural Networks have also been studied in this problem. Methods based on
backpropagation networks are presented in [7,8,9], the use of Support Vector
Machines is proposed in [10], and the Pulse Coupled Neural Network in [11].
S. Omatu et al. (Eds.): IWANN 2009, Part II, LNCS 5518, pp. 1248–1255, 2009.
c
Springer-Verlag Berlin Heidelberg 2009