7/29/2019 Dsp Numberplate Rec
1/32
VEHICLE NUMBERPLATE RECOGNITION SYSTEM
7/29/2019 Dsp Numberplate Rec
2/32
Information and constraints
Character recognition using moments.
Character recognition using OCR.
Signature.
Mostly implemented using hardware.
7/29/2019 Dsp Numberplate Rec
3/32
Technique used - Signature
Take binary image.
Sum of white or black pixelsin each row and column.
Finding peaks and valleys inrow histogram or columnhistogram.
Ridge in the Row
signature
Row Histogram - Signature
Column Histogram
Signature
7/29/2019 Dsp Numberplate Rec
4/32
Brief overview of the system
Takes image of the car and searches for the number plate in theimage.
Once the probable number plate area is located it is given to
OCR. If OCR doesnt recognize the characters from the image number
plate area is searched again from the image.
If characters are recognized then number plate search isterminated.
7/29/2019 Dsp Numberplate Rec
5/32
Limitations of the system
Noise free image with uniform illumination.
Numbers displayed in one line on the number plate.
7/29/2019 Dsp Numberplate Rec
6/32
Basic components of the system
Image division into small images (finding probable numberplate area in the image).
Recognizing number plate area.
Parsing number plate to extract characters. Apply OCR to the parsed characters.
7/29/2019 Dsp Numberplate Rec
7/32
Part I
Finding probableNumber Plate images
7/29/2019 Dsp Numberplate Rec
8/32
Finding probable number plate image
Using signature technique to break the vehicle image intosmaller image pieces.
One of these image pieces will be number plate.
Breaking image into pieces was the main issue.
7/29/2019 Dsp Numberplate Rec
9/32
Steps towardsrefinement
Thresh holding using average of minimum and maximum value ofthe signature.
7/29/2019 Dsp Numberplate Rec
10/32
Steps towardsrefinement Cont.
Thresh holding using average of non zero minimum and secondhighest peak of the signature.
7/29/2019 Dsp Numberplate Rec
11/32
Steps towardsrefinement Cont.
Thresh holding using average of minimum and minimum peak of thesignature.
7/29/2019 Dsp Numberplate Rec
12/32
Row wise signature of binarised image
7/29/2019 Dsp Numberplate Rec
13/32
Row wise signature of inverted binarised image
7/29/2019 Dsp Numberplate Rec
14/32
Car Images
Original Image
Binarised image
Inverted binarised image
7/29/2019 Dsp Numberplate Rec
15/32
Steps towardsrefinement Cont.
Thresh holding using average of minimum and median of thesignature.
7/29/2019 Dsp Numberplate Rec
16/32
Extracting Number plate from the image
One piece of image that will be tested for number plate
7/29/2019 Dsp Numberplate Rec
17/32
Part II
Recognition of theNumber Plate
7/29/2019 Dsp Numberplate Rec
18/32
Recognizing plate and parsing it
Looking for number plate in the broken pieces of vehicle image. Apply peak to valley to the candidate image pieces to further break
the image piece into possible character. Image piece with maximum peaks in candidate character is selected
as the number plate.
Column signature of the numberplate image
Column signature of the anotherimage piece
7/29/2019 Dsp Numberplate Rec
19/32
Recognizing plate and parsing it Cont.
Column signature of the another
image piece
7/29/2019 Dsp Numberplate Rec
20/32
Recognizing plate and parsing it Cont.
Column signature with maximum number of ridges. Percentage width of the ridge should be 15% of the whole number
plate image. Taking minimum value of the column histogram as thresh hold value.
7/29/2019 Dsp Numberplate Rec
21/32
Parsing plate
Images of all characters
7/29/2019 Dsp Numberplate Rec
22/32
Optical Character Recognizer -OCR
7/29/2019 Dsp Numberplate Rec
23/32
Recognition of Characters.
Method of recognition of characters froman image containing these characters is
based on object recognition techniquesused in Digital Image Processing.
Two commonly used techniques for objectrecognition areTemplate Matching using Correlation.
Distance Measurement.
7/29/2019 Dsp Numberplate Rec
24/32
Object Recognition Techniques.
Distance Measurement is based onrepresentation technique which usesmoments of an object.
Moments represent such measurements ofan object which can represent featuresassociated with that object, such as center ofgravity, Eccentricity.
Distance Measurement has some drawbackssuch as Extensive computations reduce efficiency of
execution of algorithm.
Difficult to implement.
7/29/2019 Dsp Numberplate Rec
25/32
Template Matching
Another technique which was used in thisproject for character recognition is templatematching using correlation.
The technique is based on performingcorrelation between segmented image fromwhich a character is required to berecognized and character template image
which is used for recognition. This technique is efficient as compared to
distance measurement.
Only problem is associated with templateimage i.e. proper acquisition of template imageis required.
7/29/2019 Dsp Numberplate Rec
26/32
Correlation
Modified form of convolution.
f(x,y) function represents gray scale value ata specific element (x,y) in an image.
f(x,y) represents an image from which acharacter is required to be recognized.
g(x,y) represents image of a charactertemplate.
h(x,y) represents result image aftercorrelation.
7/29/2019 Dsp Numberplate Rec
27/32
Correlation
Correlation in spatial domain can berepresented as:
h(x,y) = f(x,y) * g*(x,y)
Correlation in frequency domain can berepresented as:
h(x,y) = f 1{f {f(x,y)} .* {g*(x,y)}}
In MATLAB correlation in frequency domain canbe easily represented as:
h=real(ifft2(fft2(f,70,324).*fft2(rot90(g,2),70,324)));
7/29/2019 Dsp Numberplate Rec
28/32
Result of Correlation
As it is basic form of convolution, itsresult is an image which represents
convolutions of two matrices. The sizeof result matrix will be increased frominput image matrices.
Due to which we have to apply some
thresh holding on resultant image.Normally value of thresh hold is littleless than maximum value of resultant
image.
7/29/2019 Dsp Numberplate Rec
29/32
Detection of existence of atemplate image A single pixels presence provides
information about exact match of atemplate image with input segmentedimage.
An image after thresh holding providesinformation about the presence of apixel in correlation result image.
7/29/2019 Dsp Numberplate Rec
30/32
Modules implemented in
Matlab. Three modules were defined in MATLAB forcharacter recognition.
ocr_alpha (p1, p2, p3, p4);
ocr_numeric(p1,p2,p3,p4); P1 represents segmented image which is
required to be recognized.
P2,p3,p4 represent template images which will
be used for recognition Template(numplateimage, charimage);
Numplateimage represents the segmented image.
Charimage represents template image which is
currently used for recognition.
7/29/2019 Dsp Numberplate Rec
31/32
Template module returns no of whitepixels in resultant correlated image.
Template function is called byocr_numeric() and ocr_alpha() functions.
Ocr_numeric() and ocr_alpha() functions
are called by user and parameters tothese functions are passed by user
Modules implemented in
Matlab.
7/29/2019 Dsp Numberplate Rec
32/32
Thank you