RegularTy p e (2006/11/1)BeforeLarge
vehiclesRental carsPersonalCommercialPersonalCommercialAfter
(2006/11/1)335 × 170mm335 × 170mm440 × 220mm 440 × 220mm440 × 220mm 440 × 220mm335 × 170mm335 × 170mm335 × 155mm335 × 155mm520 × 110mm520 × 110mm520 × 110mmFigure 1: Different types and sizes of Korean LPs.thresholding to obtain maximum edge information from
those images. The detection results obtained in this stage are
remarkable for poor quality images. After finishing the two-
stage cascade AdaBoost, we find that the average LPD rate is
98.38% with a computational time of 49 ms.
This paper is organized as follows. Background and
challenges are illustrated inSection 2,andourproposedLPD
method is described inSection 3. The experimental results in
Section 4showthattheproposedmethodisabletoensure
fast LPD as well as achieve sufficient accuracy. Finally, the
conclusion is summarized inSection 5.
2. Background and Challenges
To properly work with LPR systems, we must manage a large
variety of LPs, especially in South Korea. Each province in
Korea has its own LP color, pattern, and formats of numbers
and other characters. Different colors represent different
types of vehicles. Moreover, there are three different sizes of
LPs available in Korea, such as large (520 mm×110 mm),
medium (440 mm×200 mm), and small (335 mm×170 mm
or 155 mm).Figure 1showsthedifferenttypesandsizesofLPs
available before and after November 01, 2006, in Korea.
3. Proposed System
Our proposed LPD system consists of two parts; the first part
uses cascade AdaBoost and the second part uses adaptive
thresholding. (SeeFigure 2for the system architecture of our
proposed system.)
3.1. Using Cascade AdaBoost.Using cascade AdaBoost for
LPD systems consists of two phases, offline training and
online detection, as shown inFigure 3.
3.1.1. Offline Training Phase.At the core of the offline training
phase are the training and combination of strong classifiers.
First, a series of weak classifiers (critical features) with their
weights are extracted after being trained by a large number
of positive and negative examples. Then, strong classifiers are
selected from the weak classifiers according to their weights.
Figure 4shows how the algorithm is structured. The strong
Input images from databaseSuccessfully
detected license
plate imagesUnsuccessful
detected license
plate imagesUsing cascade AdaBoost
Training phase
Detection phaseInput images from unsuccessful imagesUsing adaptive thresholding
Training phase
Detection phaseStage 1Stage 2Figure 2: System architecture of the proposed LPD system using
cascade AdaBoost.NoYe sLicense plate
examples
(positive samples)Background
examples
(negative samples)Normalized with
boundary paddingImage preprocessingAdaBoost algorithm
for training LPCreate a detector cascade structureInput test image or frameImage resizedImage preprocessingAdaBoost algorithm
for detecting LPDetection resultsSave license plate for
character recognitionRejectTraining phaseCCACheckDetection phaseFigure 3: Framework of the proposed LPD system using cascade
AdaBoost.Weight 1Feature 1
Weight 2Detector cascade structureWeight 1
Weight 2We i g h t n We i g h t nAdaBoost
algorithmFeature
extractionImage
preprocessingNormalized with
boundary paddingNegative
samplesPositive
samplesFeature 2Feature 1
Feature 2Featuren Featuren
.....
...
.Strong classifier 1 Strong classifier nFigure 4: Structure for the offline training phase.classifiers are then constructed in a detector cascade structure
for the online recognizing module.(1) The Training Databases. For offline training, positive
sample images and negative sample images are required. The
positive sample images are LP images only; the negative
sample images are background images without an LP image.(a) The Positive Samples. The Korean LP is made of
three different sizes, large (520 mm×110 mm), medium