1. Barcode detection and recognition using the Gabor wavelet.

2. Motivation. An omnipresent identification standard: the barcode.
->Unattended barcode recognition using a low camera resolution.
->Feature extraction using the wavelet theory.

3. Key words: Convolution filters. Gabor wavelet. Morphology.
Hough and Radon transform.
Feature extraction.
Classification.

4. Outline of the presentation,
Current barcode systems.
The 5 steps of the project:
Labelling
Grouping
Conditioning
Extracting
Matching
Evolutions: Feature extraction computed directy from the wavelet coefficient

5. Current barcode systems.
Hardware:
2 types of barcode readers:
Attended:
pencil-like:
gun-type reader.
Unattended: supermarket like.

6. Physical standards: barcode Charcter set Length application example
EAN 13
Number only
12 data
1 check sum
Retail product, world wide
Code128
ASCII
Variable
Widely used
PDF 415 93
Encoding parcel destination.

7. Encoding standard Optional checksum characters are include.
Encoding dependent on the standard:
Black and white bars: UPC-A
Widths of the bars: code 39.
 Barcode used: code 39 style.

8. Conclusion Effective method for identifying items, Cheap, Durable,
Easy to produce.

9. Proposed solution: A 5 steps methodology: Labelling Grouping
Conditioning
Extracting
Matching.

10. Detection of the ZOI and Gabor Wavelet:
Barcodes are a grating of oriented bars.
Easy localisable by human but not by machine.
solution: model of the primary visual cells located in the cortex, The Gabor Wavelet.
Hubel and Wiesel (1962)

11. The mathematical model:
The mother wavelet:
Parameters:
Theta: orientation
lambda: preferred wavelength
Gamma: eccentricity
Sigma: standard deviation, size of the visual field.

12. Space domain

13. Frequency domain. =4 pixels.

14. Frequency domain =6 pixels.

15. Gabor wavelets are not orthogonal.

16. Basis restriction

17. Execution: 2 octave.

18. Filtering
Lamda=2,3,5 pix
Threshold
Thin bars.
Wide bars.

19. Logical Union of the sub-spaces.

20. Conclusion for Labelling.
Three scales wavelet analysis on one direction.
The Gabor wavelet react to parallel oriented lines.
Not orthogonal -> not for Compression purpose
Simple operations will detect the ZOI.

21. Grouping: Connected component analysis.
Detection of the ZOI.
Change of logical unit:
From pixels,
To set of pixels.
Index -> belonging to a region.

22. Connected component analysis.1 X 1 X
What is a neighbhor?
-> Detection of small objects contrasting with the background.
A row of an image of a time, example (Haralick 1981)
Exemple: 1 1 2 A

23. Execution:
Delatiion will be explained later on.

24. Proposed solution. A 5 steps methodology: Labelling Grouping
Conditioning
Extracting
Matching

25. Conditioning: Noise removal.
Convolution filters, smoothing filters.
Local average (box filter). Ex: [ ]
Gauss filter. Ex: [ ]
Order statistic operators.
Median filter.
Morphological noise cleaning.

26. Convolution filters: Local average
Convolution for LTI systems. Study of FIR.
separability.
Space Domain.ex [1 1 1]
Frequency domain.

27. Convolution filters: gauss
No ripple.

28. Example: Gaussian noise variance=0.1.
Defocusing. Filter:1/16 [ ]
Gain=0.65 dB.

29. Order statistic filters: The median filter.
Linear filter for gaussian noise but poor for binary noise.
Linear combination of the sorted values.
K*K neighbourhood. K odd.
Median:
Intequartile:
threshold:

30. Example: Noise density= 0.1. Median filter 3*3 Gain= 10.42. dB
Ideal picture.

31. Let be X the binary picture and B the SE.
The binary morphology.
Identifying maximal connected sets of pixels participating in the same kind of events
First used by Kirsch(1957),2 basis operations.
Let be X the binary picture and B the SE.
Dilatation. (Minkowski addition)
When any point of B with origin x(i,j) are in X
Erosion. (shrink or reduce)
When all points of B with the origin x(i,j) are in X

32. Complexity. Let be a L*L binary pixels a SE of 2^M pixels.
British museum algorithm: L*L*2^M
L*L*2*M (Haralick 1986)

33. Derived operators. Opening Closing
Conndition for complet noise removal:
A close under K.
Opening with small circle -> remove salt & paper noise.
Extract and handle of a shape.

34. Example
Dilatation
Erosion. 1 1 1 1 1 1

35. Execution:

36. Proposed solution. 5 steps methodology: Labelling Grouping
Conditioning
Extracting
Matching

37. Parameter extraction. Template matching dependent of:
Noise
Rotation
Scale.
Solution: parameter extraction.
Ex: for a an elipse,
Its center
Exentriciy
Size

38. Example: Hough transform
used for pattern recognition in the 80’s.
Detect primitive shapes, like line, elipse ...
Used on binary image preprocessed with edge detection technique.
Point in space parameter domain.

39. Line detection.
Placer les figures

40. Adopted solution:
A simple linear regression was preferred.

41. Angle correction:

42. Signature segmentation analysis.

43. Proposed solution: A 5 steps methodology: Labeling Grouping
Conditioning
Extracting
Matching

44. Matching. Classification of the parameters. Sharp clusters
Fuzzy cluster
Neural network.

45. 2 cluster segmentation:
Distance between each pair of observations.
M observations,
N variables.
M*(M-1)/2 pairs.

46. Dendrogram Hierarchical tree.
Hight= distance between 2 clusters to be connected.

47. Conclusion on the advancement.
Working order:
256*256 pixels picture or 512*512.
Recognition till 64*64.

48. Proposed solution: Current barcode systems.
The 5 steps of the project:
Labelling
Grouping
Conditioning
Extracting
Matching
Evolutions: Feature extraction computed directly from the wavelet coefficient

49. Feature extraction computed from the wavelet coefficient
Present difficulties:
Which wavelet ?
Which basis?
compression VS classification
Irrelevant cost fonction minimisation .
Significant differences come from low energy subbands. Best Basis algrithm (Saito)
Which Coeficients?

50. From Coeficients to features
Direct computation.
Dimmension reduction.
Parameter selection
Prameter projection.

51. Conclusion: Large subject: Program in working order.
Wavelet
Digital filters
Classification
morphology
Program in working order.
Optimisation gabor model.
Feature extraction possible with wavelet.