Proceedings of the International Conference on VLSI and Communication Engineering, April 16th–18th 2009
Digital Image Processing Approach for Inspecting and
Interpolating Cracks in Digitized Pictures
Sachin, V. Solanki and A. R. Mahajan
Abstract: With the effective image processing tools, the digital image
processing technique can be applied to the virtual restoration of
artistic paintings. For the restoration of artistic work, digital image
processing technique serves many purposes. This paper presents
approach for virtual digital restoration of artistic paintings which
are suffering from the breaks or cracks. An approach includes
detection or identification and removal of cracks using digital image
processing technique. The cracks are identified by thresholding the
output of the top-hat transform. Afterward, the breaks which are
wrongly identified as cracks are separated using a semi-automatic
procedure based on region growing. Finally, in order to restore the
image, order statistics filters are used.
I. INTRODUCTION
The paper aims is to classify cracks into paintings to aid in
damage assessment [1]. Cracks are the breaks mainly
influenced by ageing, drying and mechanical factors. With
the help of digital image processing technique, cracks are
detected and then eliminated which provide help to artwork
historian, museum curators and normal public on how the
artwork would look like in earlier state [3-4]. The
methodology consists of the different stages including crack
detection, separation of the patterns which are misidentified
as crack and the final stage is to fill the break to virtually
restore the artwork [2].
This paper is organized into different sections. Section
II describes the crack-detection procedure. Method for the
separation of the crack like pattern is presented in Section
III. Methods for filling the cracks with image content from
neighboring pixels are proposed in Section IV.
II. IDENTIFICATION OF CRACKS
Cracks normally contain lower luminance and, hence it is
assumed as local intensity minima. Hence, a crack detection
process is operated on the image luminance component. Also
it should be capable to detect such minima. Top-hat transform
is proposed in this paper for crack detection procedure. The
top-hat transform is filter defined as
y(x) = f(x) – fnB(x)
(1)
where f nB(x) is the opening of the function f(x). The
structuring set nB, defined as
nB = B ⊕ B ⊕ B----B (n times)
(2)
Sachin, V. Solanki and A. R. Mahajan, is Post Graduate, Department of
Computer Science & Engineering, G. H. Raisoni College of Engineering,
Nagpur, India, E-mail: [email protected]
In the above equation ⊕, represents the dilation
operation. After evaluating the final structuring set once
using equation (2), it is used subsequently in the opening
operation of equation (1). The opening function is a lowpass nonlinear filters which removes all local maxima for
which the structuring element cannot adjust. Hence, the
digital picture f- fnB contains only those local maxima and
no background at all. As cracks are local minima, not local
maxima, the top-hat transform should be applied on the
inverse luminance image. The output of the detection
procedure can be controlled by selecting the suitable values
for the user can control the result of the crack-detection
procedure by choosing appropriate values for the following
quantities:
• Structuring element type;
• Structuring element size and the number of dilations
in equation (2).
These quantities produce an effect on the size of the
resultant structuring element and must be selected according
to the broadness of the cracks to be detected. The top-hat
transform produces a grayscale resultant image where pixels
with a high grey value are potential crack or crack-like
elements. Therefore, a thresholding operation on resultant
image is required to separate cracks from the rest of the
image. The number of image pixels which are separated as
cracks decreases as the threshold value increases. Hence
certain cracks, especially in dark image areas where the local
minimum condition may not be achieved, can remain
undetected. Therefore it is better to select the threshold value
so that some cracks remain undetected than to select a
threshold that would result in the detection of all cracks but
will also falsely identify as cracks, and subsequently modify,
other image structures.
III. SEPARATING BREAKS WHICH ARE
FALSELY IDENTIFIED AS CRACKS
Some artistic paintings contain certain breaks where they
have almost the same broadness and luminance features as
cracks. Therefore, the top-hat transform might misclassify
these breaks as cracks. Thus, in order to avoid any
undesirable changes to the original digital painting, it is very
important to separate these breaks from the actual cracks,
before to put into the effect of the crack filling method. A
procedure to accomplish this aim is described in the
following subsection.
&
Proceedings of the International Conference on VLSI and Communication Engineering
(A) Semi-Automatic Crack Separation
An easy user friendly technique for the separation of cracks
from breaks is to apply a region growing algorithm on the
thresholded result of the top-hat transform, starting from
pixels (seeds) on the actual cracks. The pixels are selected
by the user in real time. One seed per connected crack
element should be chosen. In the similar way, the user can
apply the method on the breaks, if this is more appropriate.
The growth mechanism examines recursively for undetected
pixels with value 1 in the 8-neighborhood of each crack pixel.
At last phase of this procedure, the pixels in the binary digital
picture, which correspond to breaks that are not 8-connected
to cracks, will be cleared. An example is shown here. Cracks
are normally considered as darker than the background and
that they are featured by a similar gray level, tracking is
performed on the basis of two important characteristics:
absolute gray level and crack similarity. At some point the
system recognizes some pixels lie on a crack, it allots to the
crack new pixels if their gray levels fall in a given range and
do not distinguish importantly from those of the pixels which
are already identified as belonging to the crack. Figure 1
show three iterative steps of the tracking process. Point A is
the starting crack point (seed) selected by the user. First of
all, the 8-neighborhood of pixel X is assumed (pixels Y1
through Y5). Now for every pixel Yi of the neighborhood,
the system checks the following conditions:
|f (X) – f (Yi)| <= D
(3)
f (Yi) ª [D1, D2]
(4)
where f(Yi) denotes the gray level of Yi, and D, D1 and D2
are thresholds measured as per the knowledge of the crack
pixels previously identified as such.
3 and 4 for each of them. More importantly, the validity of
condition 3 is checked for each pair of pixels belonging to
the similar 8-neighborhood. Thus, considering an example,
a pixel Zi is supposed to belong to the crack if, in support to
condition 4, condition 3 is checked for at least one Yj in the
neighborhood of Zi (in Figure 2, point Z5 is identified as a
crack pixel). The above process performs iterations until the
system can’t get a pixel with the appropriate features. The
best feature of the process is that it traverse cracks by fronts
({Y1, Y2, Y3} at iteration 1, {Z5} at iteration 2, and {W2,
W3} at iteration 3).
IV. INTERPOLATION TECHNIQUE
After detecting cracks and separating misidentified breaks,
the last step is to restore the digital picture using local image
information (i.e., neighborhood pixels information) to
interpolate the cracks. This paper proposes order statistics
filtering technique for this purpose. This does not produce
any effect on those pixels which do not belong to cracks.
Hence, as the classified crack pixels are indeed crack pixels;
the interpolation process does not affect the “useful and
important” content of the image.
(A) Order Statistics Filters For Interpolation
An important way to interpolate the cracks is to use median
or other order statistics filters in their neighborhood. All
filters are operated selectively on the crack, which means
the center of the window of the filter traverses only the crack
pixels. If the filter window is adequately bigger, the crack
pixels inside the window will be on the outline and will be
put aside. Thus, the crack pixel will be allotted the value of
any one of the neighborhood noncrack pixels. The different
filters can be used for this purpose.
• Median filter
yi = med (xi-v , ----- xi, ----- xi+v)
(5)
•
Figure 1: Tracking Process
By considering to Figure 1, the system observes that
only pixels Y1, Y2, and Y3 belong to the crack. The process
continues by referring all the pixels adjacent to Y1, Y2 and
Y3, named as Z1 through Z6. The system checks conditions
Recursive median filter
yi = med (yi-v , -----, yi-1, xi, ----- xi+v)
(6)
where the yi-v , -----, yi-1 are the initially computed median
resultant samples. For both the above filters, size of the
filter window (now only rectangular windows) should be
fairly 50% wider than thickest crack visible on the image.
This is important to ensure that the filter result is chosen
to be a value of the noncrack pixel. Windows having small
size will result in cracks that will not be adequately filled
whereas windows that are much thicker than the cracks will
cause large uniform areas, thus misrepresenting fine image
details.
• Weighted median filter
yi = med(w–v ◊% xi-v ,----- , wv ◊% xi+v)
(7)
where ◊ denotes repetitions of x, w times. For this type of
filter, smaller filter windows (approximately 30% wider than
the thickest crack visible on the image) can be taken since
the probability that a color value corresponding to a crack
is chosen as the filter result can be restricted by having small
Digital Image Processing Approach for Inspecting and Interpolating Cracks in Digitized Pictures
'
weights for the pixels centrally present within the window
and larger ones for the other pixels.
• A variable of the modified trimmed mean (MTM)
filter which do not include the samples in the filter
window, which are notably smaller from the local
median and averaging the remaining pixels
yij = ΣΣA αrs xi+r, j+s D ΣΣA αrs
(8)
The entire filter window is covered by the summations.
The filter coefficients are selected as follows:
αrs = 0 if med {xij} – xi+r, j+s >= q
= 1, otherwise
(9)
The actual cost of trimming depends on the positive
parameter. Data of smaller value diverging strongly from
the local median (which represent normally to cracks) are
trimmed out. Windows used with the variant of the MTM
filter may also be smaller than those taken for the median
and recursive median filters since a part of the crack pixels
is supposed to be avoided by trimming procedure.
V. EXPERIMENTAL RESULTS
In the experiments, we adopted the performance
identification of cracks. We tested the performance on some
images and asked different users for the inspection of the
images. The resultant images contain the identified cracks
of the images.
The resultant image is the binary output of the Top Hat
transform on the luminance component of the cracked image.
Following parameters were used for the experiment.
Figure 3: Resultant Image 1 of Identified Cracks
• Squared structuring element
• 5 x 5 size of structuring element
• 2 dilations
The Top Hat transform generated a gray scale resultant
image where the large gray value pixels are the actual cracks
or crack like structures.
The resultant image contained the higher gray values
as the cracks or similar structure hence, a thresholding
operation is performed on resultant image to separate cracks
from the rest of the image. Trial and Error method is used to
select threshold values and inspected the result by changing
the threshold values in real time. The resultant image
(Figure 3) has the threshold value 7 which is selected by
trial and error method.
A comparative resultant image of the original image
(figure 1) is shown in the figure 4 which is having the
threshold value 37.
Comparing the resultant images figure 3 and figure 4, it
is observed that the numbers of image pixels which are
separated as cracks decreased as the threshold value is
increased.
VI. CONCLUSION AND FUTURE WORK
Figure 2: Actual example Image
In this paper, we have presented a technique for inspection
and interpolation of cracks in digitized paintings. Cracks
are identified by using top-hat transform, whereas the
breaks, which are misclassified as cracks, are separated
by a semi-automatic approach. Crack interpolation is
performed by suitable modified order statistics filters.
Experiment is performed on images for identifying cracks
Proceedings of the International Conference on VLSI and Communication Engineering
or cracks like features and result is also presented by using
different parameters. Future work will focus on separating
breaks which are misidentified as cracks and applying
interpolation technique for the restoration of digitized
artistic pictures.
REFERENCES
[1] I. Giakoumis and I. Pitas, “Digital Restoration of Painting
Cracks,” in Proc. IEEE Int. Symp. Circuits and Systems, 4, 1998,
269–272.
[2] I. Pitas, C.Kotropoulos, N. Nikolaidis, R.Yang, and M. Gabbouj,
“Order Statistics Learning Vector Quantizer,” IEEE Trans. Image
Process, 5(6), 1048–1053, 1996.
[3] M. Bertalmio, G. Sapiro, V. Caselles, and C. Ballester, “Image
Inpainting,” in Proc. SIGGRAPH, 2000, 417–424.
[4] M. Barni, F. Bartolini, and V. Cappellini, “Image Processing for
Virtual Restoration of Artworks,” IEEE Multimedia, 7(2), 34–
37, 2000.
Figure 4: Resultant Image 2