speaker: ching-hao lai( 賴璟皓 )

22
NTIT IMD NTIT IMD 1 Speaker: Speaker: Ching-Hao Lai( Ching-Hao Lai( 賴賴賴 賴賴賴 ) ) Author: Author: Hongliang Bai, Junmin Zhu and Chang Hongliang Bai, Junmin Zhu and Chang ping Liu ping Liu Source: Source: Proceedings of IEEE on Intelligent Proceedings of IEEE on Intelligent Transportation Transportation Systems, Volume 2, O Systems, Volume 2, O ct. 12-15, 2003, ct. 12-15, 2003, P.P. 985 - 987 P.P. 985 - 987 A Fast License Plate A Fast License Plate Extraction Method on Complex Extraction Method on Complex Background Background Date: 2004/10/6

Upload: lane

Post on 14-Jan-2016

135 views

Category:

Documents


0 download

DESCRIPTION

A Fast License Plate Extraction Method on Complex Background. Author: Hongliang Bai, Junmin Zhu and Changping Liu Source: Proceedings of IEEE on Intelligent Transportation Systems, Volume 2, Oct. 12-15, 2003, P.P. 985 - 987. Speaker: Ching-Hao Lai( 賴璟皓 ). Date: 2004/10/6. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 11

Speaker:Speaker: Ching-Hao Lai( Ching-Hao Lai( 賴璟皓賴璟皓 ))

Author:Author: Hongliang Bai, Junmin Zhu and Changping Liu Hongliang Bai, Junmin Zhu and Changping Liu

Source:Source: Proceedings of IEEE on Intelligent Transportation Proceedings of IEEE on Intelligent Transportation Systems, Volume 2, Oct. 12-15, 2003, Systems, Volume 2, Oct. 12-15, 2003,

P.P. 985 - 987 P.P. 985 - 987

A Fast License Plate Extraction A Fast License Plate Extraction Method on Complex BackgroundMethod on Complex Background

Date: 2004/10/6

Page 2: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 22

Author:Author: Yanamura, Y.; Goto, M.; Nishiyama, D.; Soga, M.; Yanamura, Y.; Goto, M.; Nishiyama, D.; Soga, M.; Nakatani, H.; Saji, H.;Nakatani, H.; Saji, H.;Source:Source: Intelligent Vehicles Symposium, 2003. Intelligent Vehicles Symposium, 2003. Proceedings. IEEE , June 9-11, 2003Proceedings. IEEE , June 9-11, 2003 Pages:243 - 246Pages:243 - 246

Extraction and Tracking of the License Plate Extraction and Tracking of the License Plate Using Hough Transform and Voted Block Using Hough Transform and Voted Block

MatchingMatching

Page 3: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 33

OutlineOutline IntroductionIntroduction Overview of the proposed systemOverview of the proposed system Experimental ResultsExperimental Results ConclusionConclusion

Page 4: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 44

LPR has turned out to be an important LPR has turned out to be an important research issue.research issue.

LPR system consists of three parts:LPR system consists of three parts:

License plate detectionLicense plate detection

Character segmentationCharacter segmentation

Character recognitionCharacter recognition A fast license plate localization algorithm A fast license plate localization algorithm

for monitoring the highway ticketing for monitoring the highway ticketing system.system.

Introduction(1/2)Introduction(1/2)

Page 5: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 55

LP detect method overview:LP detect method overview:

Morphological operationsMorphological operations

Edge extractionEdge extraction

Combination of gradient featuresCombination of gradient features

Neural Network for color classificationNeural Network for color classification

Vector quantizationVector quantization

Back-propagation neural network (BPNN)Back-propagation neural network (BPNN)

Introduction(2/2)Introduction(2/2)

Page 6: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 66

Input ImageInput Image

Vertical Edge DetectionVertical Edge Detection Edge Density Map GenerationEdge Density Map Generation Binarization and DilationBinarization and Dilation License Plate LocationLicense Plate Location

Output RegionOutput Region

OverviewOverview

Page 7: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 77

Horizontal Sobel FilterHorizontal Sobel Filter

g(h)=|[f(i-1,j-1)+2f(i-1,j)+f(i-1,j+1)]g(h)=|[f(i-1,j-1)+2f(i-1,j)+f(i-1,j+1)]

-[f(i+1,j-1)+2f(i+1,j)+f(i+1,j+1)]|-[f(i+1,j-1)+2f(i+1,j)+f(i+1,j+1)]|

VerticalSobel FilterVerticalSobel Filter

g(v)=|[f(i-1,j-1)+2f(i,j-1)+f(i+1,j-1)]g(v)=|[f(i-1,j-1)+2f(i,j-1)+f(i+1,j-1)]

-[f(i-1,j+1)+2f(i,j+1)+f(i+1,j+1)]|-[f(i-1,j+1)+2f(i,j+1)+f(i+1,j+1)]|

Vertical Edge Detection(1/3)Vertical Edge Detection(1/3)

Page 8: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 88

Sobel Filter HorizontalSobel Filter Horizontal

g(h)=|(30*1+33*2+119*1)g(h)=|(30*1+33*2+119*1)

-(36*1+115*2+114*1)|=165-(36*1+115*2+114*1)|=165 Sobel Filter Vertical g(v)=|(30*1+33*2+36*Sobel Filter Vertical g(v)=|(30*1+33*2+36*

1)1)

-(119*1+115*2+114*1)|=331-(119*1+115*2+114*1)|=331

Vertical Edge Detection(2/3)Vertical Edge Detection(2/3)

Page 9: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 99

Vertical edge detector is better than Vertical edge detector is better than horizontal edge detector.horizontal edge detector.

Vertical Edge Detection(3/3)Vertical Edge Detection(3/3)

Page 10: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 1010

Density Formulation: Density Formulation: • 3 X 15 block and center at (I,j)3 X 15 block and center at (I,j)

• d(I,j) represents the edge density d(I,j) represents the edge density mapmap

Edge Density Map Edge Density Map Generation(1/2)Generation(1/2)

Page 11: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 1111

Edge Density Map Generation(2/2)Edge Density Map Generation(2/2)

Page 12: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 1212

Binarization(1/3)Binarization(1/3)

Otsu Histogram Threshold:Otsu Histogram Threshold:

Histogram-derived thresholdsHistogram-derived thresholds

Page 13: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 1313

Binarization(2/3)Binarization(2/3)

:: 變異數變異數

:: 概率 概率 (( 加權加權 )) 求 最小值求 最小值

Page 14: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 1414

Binarization(3/3)Binarization(3/3)

Page 15: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 1515

Dilation(1/4)Dilation(1/4) Before dilation, we use aBefore dilation, we use a nonlinear filter nonlinear filter

remove narrow horizontal lines.remove narrow horizontal lines.

If Bottom-Top<T (Threshold=5) thenIf Bottom-Top<T (Threshold=5) then

For(i=Top;i<=Bottom;i++) p(i)=0For(i=Top;i<=Bottom;i++) p(i)=0

Page 16: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 1616

Dilation(2/4)Dilation(2/4)

Page 17: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 1717

Dilation(3/4)Dilation(3/4) We dilate the image use a We dilate the image use a horizontal horizontal

mask.mask.

If Right-Left<T (Threshold=9) thenIf Right-Left<T (Threshold=9) then

For(i=Left;I<=Right;i++) p(i)=255For(i=Left;I<=Right;i++) p(i)=255

Page 18: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 1818

Dilation(4/4)Dilation(4/4)

Page 19: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 1919

License Plate Location(1/2)License Plate Location(1/2) Connected Component AnalysisConnected Component Analysis Feature ExtractionFeature Extraction

Aspect ratio (R) = W / HAspect ratio (R) = W / H

Area (A) = W x H Area (A) = W x H

Density (D) = N / ( W x H )Density (D) = N / ( W x H ) Combination of candidate regions by the Combination of candidate regions by the

connected densityconnected density

Getting Final Candidate regionsGetting Final Candidate regions

Page 20: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 2020

License Plate Location(2/2)License Plate Location(2/2)Blue Block Width=4 Height=6Blue Block Width=4 Height=6

Page 21: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 2121

Data Source:Data Source:

478 real scene images acquired from the real 478 real scene images acquired from the real highway ticketing stationhighway ticketing station

Resolution: 768x534Resolution: 768x534 Different Light condition:Different Light condition:

cloudy, sunny, daytime, night timecloudy, sunny, daytime, night time Different kind of vehicle:Different kind of vehicle:

van, truck, carvan, truck, car 459 of 478 (96%) image were successful detect 459 of 478 (96%) image were successful detect

100ms per image100ms per image

Experimental ResultsExperimental Results

Page 22: Speaker:  Ching-Hao Lai( 賴璟皓 )

NTIT IMDNTIT IMD 2222

ConclusionConclusion A fast license plate localization scheme is A fast license plate localization scheme is

presented in the paper.presented in the paper. The most serious shortcoming of our method is The most serious shortcoming of our method is

in falling to locate the license plate that is in falling to locate the license plate that is badly deficient.badly deficient.

It is relatively robust to variations of the It is relatively robust to variations of the lighting conditions and different kinds of lighting conditions and different kinds of vehicle.vehicle.