USING SLIDING CONCENTRIC WINDOWS FOR LICENSE PLATE SEGMENTATION AND PROCESSING C. Anagnostopoulos1, I. Anagnostopoulos2, G. Tsekouras1
G. Kouzas3, V. Loumos3, E. Kayafas3 3
School of Electrical and Computer Engineering National Technical University of Athens Athens, GREECE
[email protected] {loumos,kayafas}@cs.ntua.gr
1
Cultural Technology & Communication Dpt. University of Aegean, Mytilene, Lesvos, GREECE, 2 Information & Communication Systems Eng Dpt. University of Aegean, Karlovassi, Samos, GREECE, {canag,gtsek}@ct.aegean.gr,
[email protected]
morphology [2], color and gray-scale processing [3-5], shape recognition technique based on image algebra [6] and spatial frequency based techniques which exploit the fact that there are large variations at the license plate region. For the extraction of the plate region, techniques such as edge extraction [7], Hough transform [8 -9] vector quantization [10] and template matching techniques [11] have been applied. In addition, neural network approaches have been introduced to perform high accuracy recognition on license plate characters [12-13]. In this paper the Probabilistic Neural Network (PNN) technology has been implemented.
Abstract— In this paper, a new algorithm for vehicle license plate identification is proposed, on the basis of a novel adaptive image segmentation technique (Sliding Windows) in conjunction with a character recognition Neural Network. The algorithm was tested with 2820 natural scene gray level vehicle images of different backgrounds and ambient illumination. The camera focused on the plate, while the angle of view and the distance from the vehicle varied according to the experimental setup. The license plates properly segmented were 2719 over 2820 input images (96.4%). The Optical Character Recognition (OCR) system is a two layer Probabilistic Neural Network with topology 108-180-36, whose performance reached 97.4%. The PNN was trained to identify multi-font alphanumeric characters from car license plates based on data obtained from algorithmic image processing.
I.
II.
Hypothetically, license plates can be viewed as irregularities in the texture of the image and therefore abrupt changes in the local characteristics of the image would probably manifest the presence of a license plate.
INTRODUCTION – REVIEW OF OTHER TECHNIQUES
This paper presents a computer vision and character recognition algorithm. The focus is on the integration of a novel segmentation technique implemented in a license plate recognition system able to cope with outdoor conditions if parameterized properly. Specifically, the novel contributions are: •
A novel segmentation technique named Sliding Windows (SW) used for faster detection of Regions of Interest (RoI). In [1] the RoIs were textile defects, whilst in this application the RoIs are the license plates.
•
An algorithmic sequence handling plates of various sizes and positions. More than one license plate can be segmented in the same image.
•
Operation in different natural backgrounds, angles of vision in ambient illumination conditions.
•
A trainable OCR system based on Neural Network technology, subjective to continuous improvement.
Based on the above, this paper proposes a novel adaptive segmentation technique named Sliding Windows (SW). The method is developed in order to describe the “local” irregularity in the image using image statistics such as standard deviation and/or mean value. The algorithm was developed implementing the following steps: •
Creation of two concentric windows A and B of size X1xY1 pixels and X2xY2 respectively for the first pixel of the image (upper left corner). The windows are presented in Fig. 1a.
•
Calculation of statistical measurements in A and B.
•
Definition of a segmentation rule: If the ratio of the statistical measurements in the two windows exceeds a threshold set by the user, then the central pixel of the windows is considered to belong to a Region of Interest (RoI).
So, let x,y be the coordinates of the examined pixel in inspected image I. The pixel value in the respective coordinates x,y of the resulting image IAND is set either 0 (no RoI) or 1 (RoI) according the following equations:
Until now, there have been many different approaches for the segmentation of vehicle plates, such as mathematical
0-7803-9333-3/05/$20.00 ©2005 IEEE
SW SEGMENTATION METHOD
337
SIPS 2005
MA M ≤ T and B ≤ T ⇒ I AND ( x, y) = 0 MB MA MA MB in I1 ( x, y ) if > T or > T ⇒ I AND ( x, y ) = 1 MB MA
in I1 ( x, y) if
where, M is the measurement (mean value or standard deviation). The two windows slide until the entire image is scanned as shown in Fig. 1b. a. Original image I1
a
b. SW pseudocolor image
b
Figure 1. Windows A and B scanning the image
III.
c. SW binary image IAND
LICENSE PLATE RECOGNITION
d. Image I2
The license plate recognition sequence, which is proposed in this paper, consists of two distinct parts. The first one deals with the detection of the Region of Interest (RoI), i.e. the license plate. The second part includes geometric and image enhancement operations for the successful detection of the license plate characters along with an Artificial Neural Network which performs the Optical Character Recognition task.
TABLE I.
1
f. Plate isolation
e. The binary image I3
A. First part: license plate segmentation This part consists of pre-processing, in order to detect the Region of Interest (RoI) in ambient illumination conditions. The license plate segmentation pseudocode is indicated in Table I. Some steps are highlighted in Fig. 2.
Figure 2. Steps for license plate segmentation. a) The initial image, b) A pseudocolor image using the meauserements of the SW. White regions indicate possible RoIs, c) The result of SW segmentation technique after the application of the segmentation rule and thresholding, d) Image I2 after the image masking step, e) Binarization of image I2 , f) Plate detection
The following sections describe the binarization method proposed by Sauvola and the binary measurements implemented in our code.
PLATE SEGMENTATION PSEUDOCODE
for each pixel in image I1,
1) Binarization method of Sauvola In the adaptive thresholding method of Sauvola [14], the threshold T(x,y) for every pixel is calculated as a function of the coordinates x,y. This method adapts the threshold according to the local mean and standard deviation over a window size of bxb. Therefore, threshold T at pixel (x,y) is:
2 { segment I1 using SW method parameters: Χ1=5, Υ1=2, Χ2=10, Υ2=4 and Τ=1.2
} create IAND; calculate I2 : I 2 = I1 ∩ I AND ; // image masking 5 for each pixel in image I2, 6 { binarize I2 using Sauvola method parameters: k=0.5, R=128 and b=10 } 7 create I3; 8 retrieve binary object in I3 where: 9 { (Aspect Ratio>2) and 10 (Orientation
