Allocation of Text Characters of Automobile License Plates on the Digital Image

Ilya N. Trapeznikov Andrey L. Priorov Vladimir A. Volokhov

Yaroslavl Demidov State University

FRUCT-2014

Agenda

1. Introduction

2. Proposed Algorithms

3. Research Results for Detecting Number Plate

4. Research Results for Segmentation Symbols

5. Conclusion

2

The problem

3

The aim: Development an affective automobile license plate detection and number

segmentation system

4

Introduction

Conditions to the algorithms:

− descriptors in technical terms

− should not depend on priori information

− adopted to informational content on the plate

The tasks: − Design the license plate detection on digital image algorithm

− Development the symbols of the plate segmentation approach

− Creation the original image database for testing all considered methods

− Test and analysis mentioned algorithms

Conditions to the algorithms:

− descriptors in technical terms

− should not depend on priori information

− adopted to informational content on the plate

The tasks: − Design the license plate detection on digital image algorithm

− Development the symbols of the plate segmentation approach

− Creation the original image database for testing all considered methods

− Test and analysis mentioned algorithms

Automatic license plate recognition system

5

License plate detection

License plate segmentation

Symbols recognition in

segmented regions

Digital image

Car number by text representation

6

Proposed system

Corner key features detection

Region of interest (ROI) determination

Detectors for ROI calculation

Investigation license plate location

Characters segmentation

Harris Corner Detector

7

Weighted sum of squared differences between two regions

2( , ) ( , )( ( , ) ( , ))u v

S x y w u v I u v I u x v y

Taylor series expansion

yy

vuIx

x

vuIvuIyvxuI

),(),(),(),(

2

( , ) ( , )( , ) ( , )

u v

I u v I u vS x y w u v x y

x y

Matrix representation

2

2

2

2

),(yyx

yxx

u v yyx

yxx

III

III

III

IIIvuwM

Corner response function

22 β)(αkβ)(α)(tracek)det(R MM

Key points

8

α ≈ 0 β ≈ 0 key features absence

α ≈ 0 и β >> 0 edge of the object

α >> 0 и β>> 0 corner key feature

Response map of Harris corner detector

9

10

ROI construction

Corner features finding

Otsu binarization

Region merging

ROI on initial image

Local binarization

Regions binarization and merging

11

ROI on Initial Image

12

Anomaly Detection

13

Typical cases of anomaly

HOG descriptor investigtion

14

Testing database

15

Street traffic

Checkpoint

ROC-curve

16

ROC-curve is a native representation of binary classification

Research results of number plate detector

17

0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.10.89

0.9

0.91

0.92

0.93

0.94

0.95

0.96

0.97

0.98

1-Sp

Se

SI=12

SI=7

SI=8

ROC-curve for window size of corner detector

ROC-curve for binarization parameters

Research results of number plate detector

18

by number of bins in HOG histogram by scale of training images

Step 43 Step 18

19

Symbols segmentation

Detected number plate

Gradient calculation

Energy map calculation

y

Ib

x

IaIe

)(

20

Symbols segmentation

Accumulated energy at last line

Line construction rule

Research results of symbols segmentation

21

Dependence type I and type II errors on gradient parameters

Conclusions

22

Automobile license plate system was developed

Detection of license plate is carried out without making any a priori information into the system

Detection accuracy is over 97%

Other 3% occur due to low contrast and sharpness of the original image and can be changed by image preprocessing

Conclusions

23

The symbols segmentation is independed on the informational content

Optimization of the energy function increases the segmentation quality on over 98%

Allocation of Text Characters of Automobile License Plates on the Digital Image

Ilya N. Trapeznikov Andrey L. Priorov Vladimir A. Volokhov

Yaroslavl Demidov State University

FRUCT-2014