Structural Image Analysis in Investigation of Concrete

1
Classification of defectson the surface of black
ceramicsLeszek Chmielewski, Mariusz
Nieniewski,Marek Sklodowski, Waldemar
CudnyDivision of Vision and Measurement Systems
(PSWiP)Institute of Fundamental Technological
Reserach, PAS(IPPT PAN)Adam JózwikInstitute
of Biocybernetics and Biomedical Engineering,
PAS(IBIB PAN)
2
Schedule

• Objects and their defects
• Detection of defects
• Classification of defects
• Training of the classifier
• Postprocessing
• Performance of the processes
• Results
• Acknowledgements
• This research was partly supported by the
  European Commission
• ? COPERNICUS grant CRASH no. COP - 94 00717
  (1995-96)
• ? INCO-COPERNICUS grant SQUASH no. ERBIC 15CT 96
  0742 (1997-98)

3
This is not concrete ? this is ferrite

• Black ceramics
• ferrite cores
• magnets
• The material is
• milled
• molded
• pressed
• sintered
• ground
• transported
• ...
• A large number of various defects can emerge
  during these processes

A pair of ferrite cores
4
Objects and their defects called nicely
irregularities

• Surfaces3 important types of defects
• crack
• chip
• pull-out

• Sometimes difficult to classify even for humans
• Tiring quality inspection

5
Special illumination controlled light
LED illuminator
Fluorescent illuminator

• Tangential, multidirectional light amplifies the
  visibility of defects
• Brightness uniform and independent on distance

6
Locating the object in the field of view
original ? thresholded ?
complemented

• Simple morphological operations help to find the
  region occupied by the object for further
  processing
• Aim to eliminate bright spots and blobs
• ? In this application a narrow stripe at the edge
  was excluded from analysis

original ? complemented
7
Detection of irregularities (not defects!)
original ? thresholded ? elongated
irregular ? summed
Region of interest is further limited to the
irregular part of the surface with the
morphological methods ? tomorrows
presentation by prof. Mariusz Nieniewski
8
Classification of irregularities features (1/2)

• Each pixel detected in the detection phase is
  classified with the pattern recognition methods.
  Pixel pattern.
• Features are calculated for each pixel functions
  on pixel neighbourhoods masks.
• Direction invariance of features is obtained by
  rotation of the mask according to local
  directionality of texture.

original mask
rotated mask
Pixel its mask
YBF95
9
Classification of irregularities features (2/2)

• brightnesses in the original rotated mask
• statistical moments of order up to R in masks
• gradient modulus
• 9 classical textural features according to
  Law80, Pra91
• textural features based on coocurrence relations
  WuCh92
• relative values of brightness function along the
  red line
• From 30 to 150 features were used for feature
  selection.

For example Features as in Law80, Pra91
convolve the mask with A1 A9 and take standard
deviation of the output values ? 9 features.
10
Classification the K Nearest Neighbour (k-NN)
method
11
K-NN Enhancements and speed-ups

• With full selection of features and K
• Leave-one-out method
• Fuzzy version
• Fuzzy decisions made crisp in the end
• Parallel
• Distinct classifiers for each pair of classes
• Hierarchical
• Advanced version only where classes overlap
• Reference set largely reduced
• with the modified, bidirectional Hart algorithm
• ? Optimized, low error rate, quick
  algorithm

12
The parallel version of the K Nearest Neighbour
method
13
K-NN class overlap as the training criterion
(not error)
14
Training the training patterns

• Note artificial, boundary classes introduced ?
  better accuracy

2479 training patterns can be obtained quite
quickly...
15
Training first results
2479 training patterns raw classifi-cation
results enhanced by local votong
The system has successfully classified thousands
of unknown pixels. Quite satisfactory results can
be obtained with just 4 training images.
16
Training first results zoom results of the
enhancement
pixels used in training
classified pixels all / raw / enhanced
This was only a convincing example. The error
rate estimated with the leave-one-out method was
3.3. More training patterns were used in the
final system.
17
Final training results error estimates

• 5903 training patterns
• a posteriori error probabilities pixel
  classified as class "i" (row) comes in fact from
  the class "j" (column) in
• overall error 2.56
• max error 9 between classes 8 and 9
• cared for by the postprocessing (to some extent)

18
Classification results various types of
elements (1/4)
blue chip, yellow good object
19
Classification results various types of
elements (2/4)
brown irregular, red crack, green
pull-out, blue chip, grey good object.
20
Classification results various types of
elements (3/4)
blue chip, navy chip near good, red crack.
21
Classification results various types of
elements (4/4)
blue chip, green pull-out, red crack !?
22
Higher level discern cracks from grinding
grooves
Classify cracks red betweenall irregular
regions green. Limits of the method reached.
details in other images
23
Classification results rotation (in)variance
24
Technical data Performance

• Resolution
• A 512512 pixel camera. Spatial resolutions ?
  0.05mm/pixel
• (up to 20larger magnifications can be attained
  with normal lenses)
• Accuracy of results
• Classification errors overall up to 4,
  inter-class typically 4, max 10
• Stability of results
• Detection phase repeatability not worse than
  2-5 in area.
• Classification phase repeatability not worse
  than 10-20 in area, depending of how fast
  classifier version is used.
• Processing time s (PC, 1000 MHz)

only software morphological processor
image acquisition 0.05-0.20 0.05-0.20
detection 2.00 0.0001-0.001
classification 1.00 typically 10 for v. large defects - 20 of object 1.00 typically 10 for v. large defects - 20 of object
decision 0.1 0.1
25
Conclusion

• Irregularities of flat surface in black ceramics
  ferrite cores, magnets can be detected and
  classififed
• Special lighting system has been designed
• Detection of irregularities
• Irregularities in general dynamic thresholding
• Compact irregularities morphological method
• Elongated irregularities morphological method
• General decision on quality of the tested object
• Classification of irregularities
• Training by showing examples
• Segmentation and measurements
• Detailed, quantitative final decision
• Project www site http//www.tpo.org.pl/squash

26
References

• A. Józwik, L. Chmielewski, M. Sklodowski and W.
  Cudny. A proposition of the new feature space and
  its use to construction of a fast minimum
  distance classifier. In Proc. 2nd Polish
  Conference on Computer Pattern Recognition
  Systems KOSYR 2001, pages 381-386, Milków,
  Poland, May 28-31, 2001.
• M. Nieniewski, L. Chmielewski, A. Józwik and M.
  Sklodowski. Morphological detection and
  feature-based classification of cracked regions
  in ferrites. MGV, 8(4)699-712, 1999.
• A. Józwik, L. Chmielewski, M. Sklodowski and W.
  Cudny. Class overlap rate as a design criterion
  for parallel Nearest Neighbour classifier. In
  Proc. 1st Polish Conference on Computer Pattern
  Recognition Systems KOSYR'99 Trzebieszowice,
  Poland, May 24-27, 1999.
• A. Józwik, L. Chmielewski, M. Sklodowski and W.
  Cudny. A parallel net of (1-NN, k-NN) classifiers
  for optical inspection of surface defects in
  ferrites. MGV, 7(1-2)99-112, 1998.
• G. Vernazza, M. Lugg, T. Postupolski, A. Józwik,
  L. Chmielewski, D. Chetverikov and M. Peri.
  SQUASH Standard Compliant Quality Control System
  for High-Level Ceramic Material Manufacturing. In
  Proc. INCO-COPERNICUS-INTAS Workshop on Advanced
  Ceramics and Alloys, pages 35-40, Brussels,
  Belgium, Mar 12-13, 1998. European Commission,
  Directoriate Generale XII.
• L. Chmielewski, M. Sklodowski, W. Cudny, M.
  Nieniewski and A. Józwik. Optical system for
  detection and classification of surface defects
  in ferrites. In Proc. 3rd Symp. Image Processing
  Techniques (TPO'97), pages 1-13, Serock, Poland,
  Oct 29-31, 1997. Oficyna Wydawnicza Politechniki
  Warszawskiej.
• M. Mari, C. Dambra, D. Chetverikov, J. Verestoy,
  A. Józwik, M. Nieniewski, M. Sklodowski, L.
  Chmielewski, W. Cudny and M. Lugg. The CRASH
  Project Defect Detection and Classification in
  Ferrite Cores. In A. Del Bimbo, editor, Proc. 9th
  Int. Conf. Image Analysis and Processing, number
  1310 in Lecture Notes in Computer Science, pages
  781-787 (vol. II), Florence, Italy, Sept 17-19,
  1997. Springer Verlag, Berlin.
• A. Józwik, L. Chmielewski, W. Cudny and M.
  Sklodowski. A 1-NN preclassifier for fuzzy k-NN
  rule. In Proc. 13th Int. Conf. Pattern
  Recognition, pages D-234 - D-238, Wien, Austria,
  Aug 25-29, 1996. IAPR, Technical Univ. Vienna.
• Law80 K. I. Laws, Textured image segmentation,
  Univ. of Southern California, Image Processing
  Institute, USCIPI Report 940, Jan 1980
• Pra91 W. K. Pratt, Digital Image Processing,
  John Wiley, New York 1991.
• WuCh92 C-M. Wu, Y-C. Chen, Statistical feature
  matrix for texture analysis, CVGIP Graphical
  Models and Image Processing, 54, 5, 1992,
  407-419.
• YBF95 G. Z. Yang, P. Burger, D. N. Firmin, S.
  R. Underwood, Structure Adaptive Anisotropic
  Filtering for Magnetic Resonance Image
  Enhancement, Proc. 6th Int. Conf. CAIP, Prague,
  Czech Republic, Sept. 6-8, 1995, 384-391. Lecture
  Notes on Computer Science. Springer Verlag, 1995.