Vous êtes sur la page 1sur 6

SURFACE DEFECT DETECTION ON FIXED PATTERN CERAMIC TILES

F. Acebrón, F. López, J.M. Valiente, J. R. Navarro


Computer Vision Group
Polytechnic University of Valencia
Camino de Vera 14, 46022 Valencia (Spain)

ABSTRACT the most important quality control tasks to be automated.


But this is a complex task due to the diversity of surface
Automatic and reliable surface defect detection is an faults and the large variety of decorative surface patterns
important task in the ceramic tile industries area. This that can be made on production lines (plain, grained,
question has not yet been solved in a satisfactory form, so textured, ...etc). Furthermore, the patterns can be fixed,
it is already open to improve. The present work introduces random or pseudo-random textured.
a method for a particular case of ceramic tile pieces, the
fixed pattern ceramic tiles, where the production process Recent works reported in the literature approach the tile
assures that in all the pieces the same decorative surface is inspection process as a colour grading problem.
reproduced. This method is based on ‘compare with a Boukouvalas et al. [1] propose using colour histograms to
reference’ techniques. One tile without defects is used as perform the colour shape grading and study the criterion
a reference to find defects on the inspected pieces. The to measure the similarity, or dissimilarity, between colour
method is divided in to four phases: registration, histograms. Colour texture comparison has also been
difference map computation, feature extraction and proposed for the same purpose [2],[3]. In [4], Smith and
filtering, and classification. Stamp introduce a novel method to inspect curved
textured tiles.
The registration phase is necessary to spatially align both
tiles in order to be able to compare them. An accurate However, most tile production is based on decorative
registration method based on edge detection techniques is patterns with regular or irregular textures that are
presented. The feature extraction is obtained from a deterministically produced. The most common technique
difference map computed using spatial and perceptual used to detect faults on this kind of tile model is based on
colour comparison techniques. This difference map is direct comparison, on a pixel-by-pixel basis, between the
submitted to a filtering stage to eliminate tiny objects due tile image and a reference template pattern. The present
to noise induced at the acquisition and registration stages. work approach the inspection of fixed pattern tile models
Finally, one rule-based classifier is presented. This using this methodology.
classifier is oriented to detect the defects, not to quantify
them. This is useful for detecting faulty series of pieces at Our proposal is based on a four stages process. A fast
the intermediate stages of the production lines. registration stage performs the alignment of reference and
test tiles. Then a difference map is obtained computing
KEY WORDS the colour distances between both images in CIELuv and
Quality control, tile inspection, image registration, surface CIELab spaces. The faults detected on the tile will be
defect detection. determined by extracting features from the difference map
and finally they are classified using an rule-based
1. INTRODUCTION classifier. This process should be achieved in the smallest
possible time, in order to be able to adapt the inspection
In the ceramic tile manufacturing industry, a reliable and system to the production line ratios that work at 60 tiles
automated quality control process has become a necessary per minute. Therefore, minimal computer time is one of
tool for competitive purposes. In recent years these the major goals.
industries have made great efforts to automate all the
processes in the production chain. This has been achieved 2. REGISTRATION
in the majority of the processes and now automatic
quality control is the aim, not only for the final product Assuming that one or several “fault-free” tile images have
but also for the intermediate stages. previously been acquired, the first stage of tile inspection
is registration. Some methods specifically proposed for
The great majority of tile faults on the production lines tile registration [5] have been reported. Our proposal is an
are surface defects, therefore, surface inspection is one of improved version of a previously developed method [6]
based on fast edge detection techniques. Our method uses Figure 1 shows the progressive approximation of the
tile border pixels to adjust the four straight lines that fitted line to the correct edge points, for the left side.
compose the tile rectangle. These border points are After three steps, the furthest edge points are discharged
obtained by separating the tile from the image background and the correct line fitting is produced using the
using an optimal threshold level. The algorithm selects remaining points.
the histogram threshold that minimizes the cross entropy o X
between the thresholded image and the original image [7]. 3 rd
1st
Given a grey level image f(x,y), and a threshold level ‘t’, a 2n d
> σX1
bi-value image g(x,y) with values g i ∈{η1,η2} can be
computed as follows:

fi <t
fi ∑ fi
o u tliers

η1 (t ) = η 2 (t ) = f ≥t
i

> σ X2
N1 , N2 Y
where fi is the histogram value for the i level in the
original image. The cross entropy between both images e d g e p o in ts
can be computed as: Y

 f   f 
η (t ) = ∑ f i log i  + ∑ f i log i 
f i <t  η1 (t )  fi ≥t  η 2 (t ) 

The optimum threshold level is selected by:


t0 = min (η (t )) Figure 1. Line fitting process. Two outliers are discharged in three
iterations.
t
Using NP evenly spaced horizontal and vertical scan lines,
the outermost points of each tile border are selected from Figures 2 and 3 show the results obtained after applying
the thresholded image, creating four sets of NP lateral this procedure on two different tile models. Some points
edge points (PUP, PDOWN, PLEFT, PRIGTH). For each set of lying outside the tile area can be seen.
lateral points, a straight line is fitted using least-squares
method.

Let PUP={p0(x0,y0), ….., pN-1(xN-1,yN-1)} be the set of NP


edge points of the upper side. A line y = aUP + bUP x can
be fitted using the standard formulation:
σ X2 σ X2
y = aUP + bUP ⋅ x = E ( y ) − E ( x ) + 2 ⋅x
σ XY
2
σ XY
Proceeding in a similar way, the remaining lateral lines
can be computed (aDOWN,bDOWN), (aLEFT,bLEFT), (aRIGTH,
bRIGTH). During this stage, an outlier detection test is
applied to eliminate the outlying points that sometimes
appear due to the background noise or border tile defects.
Those points lying further than one and a half standard
deviations (k=1.5) from the computed line are considered
outliers, and discharged the line parameters being
computed again. The fitting process is repeated, usually
Figure 2. Lateral points and resulting tile rectangle. Due to dark
two or three iterations, until all the remaining points fulfil appearance of the tile model, a low dynamic range and poorly
the condition. This process works as follows: contrasted image is obtained. As result some details of the conveyor
belts appear, producing outlier points at the bottom side. These point
Repeat are discharged during line fitting computation.
• From the point set PUP computes the line
parameters aUP, bUP. Once the bounding rectangle is obtained, the four corner
• For each point pi(xi,yi) computes its distance to co-ordinates are computed and the inverse geometric
the line: mapping is applied as follows:
Di = | yi – ( aUP + bUP xi ) |  xir   cosα − sin α   xi   d x 
If Di <= k σ2y  r  = S ⋅  ⋅  +  
y 
 i  sin α cosα   yi   d y 
then do nothing
else extract pi(xi,yi) from PUP where p(xi,yi) and pr(xir,yir) i = 1..4 are the test and
Until (no point is extracted) reference tile corners, dx and dy are the displacements
between the top left points in the two tiles, α is the For the first question we have established that one
rotation angle, and S is the scale factor. (O,X,Y) is the difference appears when it is visible from the point of
image reference coordinate system. view of the human perception, that is just like factory
operators work. The perceptible or noticeable difference
question was firstly studied using grey levels images. The
human eye can perceive up to 40 different grey tones.
With 256 grey levels, the perceptual human difference
arises using a threshold near to 7 (256/40=6,4). However,
the results obtained using grey levels have not been
satisfactory, not detecting defects that was clearly visible.
This is because the human vision is a colour perception
system.

Our acquisition system provides RGB colour images.


Colour differences can be computed using standard
Euclidean distance. However, the RGB colour space is
not suitable to be used as a human vision related system,
because is not a perceptually uniform space [8], that is,
two closed colours may be separated by a long distance
and two visually different colours may be very close in
Figure 3. Lateral points and rectangle obtained with bright tiles. Due this space. So, we have used the CIE-Lab and CIE-Luv
to the high dynamic range and contrast of the image, background
details disappear and practically not outlier points are detected.
colour spaces, which are perceptually uniform, and the
Euclidean distance as a measure of colour difference.
It is possible to determine the parameters of the ∆E (i, j ) Lab = ( Li − L j ) 2 + (ai − a j ) 2 + (bi − b j ) 2
transformation mapping by minimizing the sum of the
squared errors: ∆E (i, j ) Luv = ( Li − L j ) 2 + (ui − u j ) 2 + (vi − v j ) 2
[ ]
4
E = ∑ xi r − S ( xi cos α − yi sinα ) − dx
2
+ Other colour difference measures can be also used in
i =1 these spaces [9], like CMC(l:c), BFD(l:c), CIE94 and
[y i
r
− S ( xi sinα + yi cos α ) − dy ] 2 CIE97. These measures are appropriated for extremely
low colour differences, which is not our case.
Replacing (S cos α) by ‘a’ and (S sin α) by ‘b’: The second question is pixel to pixel vs. pixel to area
[ ]
4
E = ∑ x − S (a ⋅ xi − b ⋅ yi ) − dx
2 comparison. The former is the standard way to perform
r
i +
i =1
image comparisons. The pixel-to-pixel distance is

[y ]
computed and a difference map is constructed when this
− S (b ⋅ xi + a ⋅ yi ) − dy
r 2
i distance is greater or equal than one given threshold. This
method does not consider the introduced ‘false
Deriving E with respect to a, b, dx, and dy and equaling differences’, due to several circumstances, like size
to zero, the following linear system of equations appears: differences, induced by tolerances during the
manufacturing or changing ambient conditions during the
∑ (xi 2 + yi 2 ) 0 ∑ x ∑ y 
i i ∑(xi r xi + yi r yi ) acquisition process. The interpolation techniques used for
i a  i  registering can also introduce pixel and/or sub pixel
∑(x + y ) − ∑y ∑x   b  ∑( yi xi − xi yi )
i i
 0 i
2
i
2
i i
r r
differences. In order to overcome this problem, one pixel
 i i  i  = i 
 ∑ xi −∑y i n 0  dx  ∑ xi r  to area technique has been used. This consist of compare
 i i     i  de test pixel with its corresponding area at the reference
 ∑ yi ∑x i 0 n    
d y
∑ yi r 
 i i   i 
image. This area is a 3x3 window established around the
reference pixel. If the minimum distance, computed with
from which the optimal parameters of the geometrical all pixels belonging to the reference window, is greater or
transformation can be obtained. This approach saves a equal than a threshold ∆ELab, the test pixel is considered a
‘real difference’.
{ }
great deal of computing time as it directly calculates the
exact inverse transformation. 1 ∆E (i, j ) = min ∆E i j ≥ ∆E Lab
t= j∈ W

0 other case
3. DIFFERENCE MAP COMPUTATION where t(x,y) is the difference image, i(x,y) the test image,
j(x,y) is the reference image and W is a 3x3 window.
Once the images have been registered, a difference map
has to be extracted comparing both images. Two Figure 4 shows the differences map obtained in both
questions arise: grey level vs. colour based comparison, cases.
and pixel to pixel vs. pixel to area comparison.
eliminates those objects accomplishing the following
conditions:
• Size filter: Area < 10 pixels (about 0.5 mm2).
• Elongation filter: PerimeterVSArea >= 0.65 and
elongation < 0.2 and no pixels touching the piece
boundary.
• Lateral effect filter: PerimeterVSArea > 0.98 and
pixels touching the piece boundary.
The size filter is needed to remove small objects due to
the noise introduced in the acquisition and registration
stages and also small defects that humans can not clearly
distinguish. Tiny displacements of the silk-screening
masks produces very thin and elongated lines in the
difference map which are not considered as defect by the
factory experts. These objects are eliminated by the
elongation filter. Finally the lateral effect filter permits to
eliminate very elongated lines located at the piece sides,
therefore these objects are produced in the registration
stage due to the tiny size variations between different
pieces. Figure 5 shows the result of applying these filters.

Figure 4. (Up) Tiles showing typical integrity and spot defects.


(Down) differences map obtained using pixel-to-pixel (left) and pixel-
to- neighbourhood window(right) distances.

4. FEATURE EXTRACTION AND


FILTERING
Once the difference map is obtained, the next step is to
study this map to detect and classify the surface tile
defects. Initially, every object in the difference map is Figure 5. The left image shows the result obtained after applying the
size filter to bottom- right image of figure 4. The right image shows the
considered as a possible defect. To characterize them a set result obtained after applying the three filters to the same image. The
of features are extracted: false defects disappear remaining the real ones.
Perimeter
Perimeter VS Area = 5. CLASSIFICATION
Area

((µ 20 + µ 02 ) + ( µ 20 − µ02 ) 2 + 4 ∗ µ11 )


2 The last stage of tile inspection is defect classification.
MajorAxisLength = 4 ∗ Several classification criteria can be proposed, depending
2 ∗ µ00
on the kind of inspection to perform. Two general cases
can be considered:
(( µ 20 + µ 02 ) − ( µ 20 − µ02 ) 2 + 4 ∗ µ11 )
2

MinorAxisLength = 4 ∗ • Final product inspection. In this case it is important


2 ∗ µ 00 to detect as many defects types as possible, and
Area quantify them using quality measurements. The goal
Area VS AxesArea = is to obtain a measure of the global tile quality that
| MajorAxis | * | MinorAxis | allow to classify it according with the standard
MinorAxisLength categories used in the ceramic industry: firs, second
Elongation = and third class tiles. The main issue is to avoid
MayorAxisLength
delivering low quality tile to exigent customers.
• Intermediate product inspection. The main goal in
A subsequent filtering stage is compulsory in order to this case is the detection of repetitive defects but not
reduce the false defects introduced by the previous stages. their quantitative measurement. This is because large
At the same time, the factory requirements indicates that series of consecutive faults are symptom of
some little objects should not be considered as faults. malfunction on the production line. Only a minor
Then, several shape and size filters are applied which
defect classification scheme should be attained to In the context of tile production, the usual colour gamut
indicate the probable source of the error and to of pieces in intermediate stages of production, that is
eliminate highly defective tiles (recyclable material). before the kiln input, are restricted to clear and pastel
shades. Then, we plan an experiment to find out suitable
We approached the later case. Basing on the experience of ∆ELab and ∆ELuv differences for these cases We artificially
factory operators, a three class-classifier is proposed create eight panels of 100 colour samples. Each panel was
capable to identify three kind of defects: obtained from the CIELuv and CIELab spaces, tracing a
• Integrity defects, produced by misalignment of straight line from the origin toward a specific colour
conveyor belts or positioning devices along the region, like yellows, pinks, green, etc. An initial colour
production line. was chosen at distance E=80 from the origin, to skip the
• Spot defects, which are spot and marks produced by dark region, and 100 colour samples were taken at
machine dripping . ∆E=0,1 intervals.
• Silk-screening defects, whose main source is the A human observer inspected each of the 8 panels and
screen misalignment or dirtiness in the silk-screening decided the targets at which a colour change were
machines. noticeable with respect to the initial colour. The
experiment was repeated with ten observers and the mean
No complex stochastic neither geometric classifier are noticeable colour differences were obtained. The results
needed in this case.. A close relation among the calculated are shown in figure 6.
features and the above defect types can be established as
follow: Mean noticeable colour differences
Integrity defects are characterized by high compactness
and usually located at the tile boundaries. The rule to 2,4 CIELab CIELuv
obtain them is: 2
Number of border pixels > 0
Area vs. Axes area > TI 1,6
Silk-screening defects have a low level of compactness 1,2
and are usually located inside the tile (out of the border).
Number of border pixels = 0 0,8
Area vs. Axes area < TS 0,4
Spot defects are very compact blobs located inside the
pieces. 0
Red Pink - Yellow- Green Green- Soft Blue Violet
Number of border pixels = 0 Orange green Blue blue
Area vs. Axes area > TM

This simple and effective rule-based classifier is enough Figure 6. Mean noticeable colour differences obtained in eight
regions of CIELab and CIELuv colour spaces.
to categorizing the defects. A subsequent counting
process can be employed to detect defect series.
It is remarkable the better human sensitivity to green
colours but, in general, there were not colours gamut
6. EXPERIMENTS AND RESULTS where minimum noticeable colour differences higher than
2 appears. These result confirm us the initial guest value
Several experiments were carried out to tune the of 2,3 for the generation of colour difference maps.
parameters involved in the algorithms and to obtain the
global performance in laboratory conditions. The next experiments try to obtain the goodness of the
proposed methodology under real conditions. For this
The first experiment tried to obtain the minimum purpose, a laboratory prototype was developed. The
perceptive difference between two colours in CIELab and prototype is composed by two main parts: the acquisition
CIELuv spaces, needed to compute the difference map module and the processing module (see figure 7). The
between reference and test tile images. It is known that acquisition module contains a colour line scan camera
CIE spaces are isotropic, even though from the point of (Dalsa Trillium), two presence photocells and a high
view of colour differences they are not completely frequency advanced fluorescent illumination system
uniform. It has been reported [5] that noticeable which assures an uniform lighting of the pieces. All these
differences between similar colours lightly change in elements are included in a dark cabin that isolates the
different regions of the colour space. However, as a piece that is been imaged from the external ambient
general rule for the CIELab space, ∆ELab differences of conditions. The processing module is composed by a
about 2,3 or greater are noticeable and those over 10 are cluster of four industrial processor boards connected
so different that comparison is not worthwhile. Then, an using a Fast Ethernet network. This module is intended to
initial value of 2,3 should be a good selection in our case. analyse all the pieces in the production line in real-time.
CABIN 7. CONCLUSIONS
An automated tile inspection system has been presented.
The proposed methodology to perform surface defect
COLOR LINE detection on fixed pattern ceramic tiles has been detailed,
LLS SCAN CAMERA including the registration, difference map, feature
CE (Trillium)
TO
P HO extraction and classification stages.
2
1
The global performance of the inspection process is about
100% of defect detection and near 99% of success in
LAMPS defect classification. Some degree of misclassification has
TILE High Frec. been introduced when, for example, spot defects located
CONVEYOR TILE
INPUT Fluorescent at the boundary of the piece were initially labelled as an
BELTS
integrity defect. These misclassifications are not
LIGHTS important from the point of view of the production
ALARM
process, where what is important is defect detection. From
this point of view, all the defects were correctly detected
during the difference map computation and most of false
POSITIONING defects were properly eliminated. These preliminary
DEVICE
ENCODER results can’t be extrapolated due the reduced number of
pieces and tile models used in the experiments. However,
the proposed methodology seems to be appropriate for the
problem of tile inspection in intermediate production
stages. Further works can be attained to improve the
system by adding new capacities for defect detection,
including tonality defects in the colours of tile model.
Figure 7. Elements of the Tile Inspection prototype. (Up) Acquisition
module details. (Down) Processing module and auxiliary devices. REFERENCES

Using the prototype an image dataset was constructed, [1] Boukouvalas, C.; Kittler, J.; Marik, R. and Petrou, M.
composed by nine tile models representative of the ”Color grading of randomly textured ceramic tiles using
factory production. For every tile model, three series of color histograms”. IEEE Trans. On Industrial Electronics.
50 pieces with integrity, silk-screening and spot defects Vol. 46, No. 1, 1999, pp. 219-226.
and a set of 10 fault-free pieces were imaged. These [2] Lumbreras, F.; Baldrich, R. et al. “Multiresolution
pieces were initially labelled by experimented human colour texture representations for tile classification”. Proc.
operators. The dataset was arbitrarily divided into training VII Nat. Symp. on Pattern Recognition and Image
and test sets (30%-70%). Analysis Vol. 1, Spain 1997, pp.145-152.
[3] Baldrich, R.; Vanrell, M. and Villanueva, J.J.
A first experiment was carried out to probe the feasibility “Texture-colour features for tile classification”.
of the registering and error map stages using the dataset. EUROPTO Conf. on Polarization and Colour Techniques
The results showed that 100% of real defects were in Ind. Inspection. SPIE Vol. 3826, 1999, pp. 124-135.
detected, but with some false defects not filtered out. [4] Smith, M.L. and Stamp, R.J. “Automated inspection
These false defects are not repeated in the tiles that of textured ceramic tiles”. Computers in Industry. Ed.
compose the series and can be removed in a post- Elsevier, Vol. 43, 2000, pp. 73-82.
processing stage. [5] Costa, C.E. and Petrou M., Automatic registration of
ceramic tiles for the purpose of fault detection, Machine
A second experiment was made with the goal of defining Vision and Applications, 2000, vol. 11, 225-230.
the thresholds values TI, TS and TM proposed in the [6] Valiente J.M.; López; F.; Acebrón, F. and Pérez E.,
classifier and compute the success rate of this stage. Due An Image Registration Method for Ceramic Tile
the simplicity of the proposed rule-based classification Inspection Purposes. Proc. of the 2001 Int. Conf. Quality
scheme, there was not necessary to make complex Control by Artificial Vision, France 2001.Vol. 2,.498-503.
computations. Only a simple selection of the minimum [7] Li,C.H. and Lee, C.K., Minimum cross entropy thres-
value to correctly classify all the training set was holding, Pattern Recognition, 1993, vol. 26(4), 617-625.
computed. Surprisingly, the value of TI, TS and TM = 0,4 [8] Stephen J. & Robin E.N., colour image processing
was appropriate for all thresholds. Afterwards, the handbook, Ed. Chapman & Hall, 1998
classification of the test set was accomplished, yielding a [9] Mahy M.L., Van Eyckdenm A., Oosterlinck,
result near to 99% of success (1% of false detection and Evaluation of Uniform Color Spaces developed after the
misclassification). Adoption of CIELAB and CIELUV, Color Research and
Appl., vol 19(2), 1994, 105-121.

Vous aimerez peut-être aussi