HASSIP/DFG-SPP1114 Workshop

1
Detection of Cardboard Faults during the
Production Process

• Nataša Babacev, Marko Barjaktarovic
• University of Belgrade, Faculty of Electrical
  Engineering
• Desanka Radunovic
• University of Belgrade, Faculty of Mathematics
• Belgrade, Serbia and Montenegro
• Bremen, Germany, 23.01.-26.01.2006.

2
Introduction

• Production of cardboard in a long bolt
• Occurrence of faults and stains on the surface
• Detection and localization in real-time
• Existing algorithm using Kirsch operator
• Discrete Wavelet Transform using
• third level Haar wavelets
• Denoising and selection of
• vertical wavelet coefficients

3
Cardboard Production

• Cardboard exits the production machine
• Optoelectronic system photographs the surface
• 8001024 pixels with 256 levels of gray
• Influence of the factory lights and optic lance
• Nonuniform distribution of gray
• Preprocessing of the image in order to get almost
  uniform distribution

4
Original image of cardboard
5
Almost uniform distribution
6
Image with fault and noise
7
Existing algorithm

• Wavelets represent the optimization of an
  existing algorithm by Marko Barjaktarovic
• High level of noise
• due to short time of exposition and
• high speed of the cardboard
• Denoising is done by a filter of size 1x5 pixels
• Extracting of the edges of faults is done by
  using the modification of Sobel operator,
• i.e. Kirsch operator 5x5 matrix

8
Result of the Kirsch operator
9
Converting to binary image

• The threshold is computed from the histogram

10
Denoising using erosion

• Clearing the image of dots that
• Kirsch operator detects as an edge
• because of local variations of gray
• The value of every pixel is replaced with the
  minimum value of neighboring pixels with the
  same -coordinate and or in
  order to preserve the line faults that occur
• Two consecutive erosions are needed

11
Erosion of the line fault
12
Line faults

• For a fault in a shape of thin line the result is
  more dots with the same -coordinate
• Distribution of area of all faults objects
  on the image on -coordinate
• If the area is large for a narrow value of
• then it is considered a line fault

13
Distribution of area of a line fault
14
Optimization of the algorithm

• Denoising is done prior the edge detection
• Edge detection
• Both are done using third level Haar wavelets

15
Denoising with DWT level 3 Haar

• Soft thresholding
• Fixed form threshold t
• s median absolute deviation of detail
  coefficients of scale 1

16
Denoised image with level 3 Haar
17
Extracting the edges

• DWT using third level Haar wavelets
• Faults occur in the direction of motion, i.e.
  vertical direction
• Only vertical detail coefficients are kept

18
IDWT with vertical coefficients
19
Binary values of pixels
20
Comparing the two methods

• Denoised image with DWT is clearer then
  in the existing algorithm
• Faster then detecting edges by Kirsch operator
• Less data stored in a sparse matrix
• After IDWT image is with less noise
• Erosion is still needed after IDWT

21
Image without faults after IDWT
22
Most common example of fault
23
Denoising with DWT level 3 Haar
24
Binary image of IDWT
25
Image after two erosions of IDWT
26
Further optimization

• The algorithm using wavelets for edge detection
  is still in development
• Possible optimization
• leaving less vertical coefficients as non-zero
• The right threshold needs to be determined

27
Detection of Cardboard Faults during the
Production Process

• Nataša Babacev, Marko Barjaktarovic
• University of Belgrade, Faculty of Electrical
  Engineering
• Desanka Radunovic
• University of Belgrade, Faculty of Mathematics
• Belgrade, Serbia and Montenegro
• Bremen, Germany, 23.01.-26.01.2006.