yaroslavl demidov state universityd2%f… · 21 dependence type i and type ii errors on gradient...
TRANSCRIPT
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 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
10
ROI construction
Corner features finding
Otsu binarization
Region merging
ROI on initial image
Local binarization
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
)(
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%