CRACK MEASURING SYSTEM BASED ON HIERARCHICAL
IMAGE PROCESSING TECHNIQUE
Takeshi Doihara, Kiichi Hirono, Kazuo, Oda
Asia Air Survey Co., Ltd.
13-16 Tamura-cho, Atsugi-shi, Kanagawa-pref., 243, Japan
Takashi Kimura, Yasuhiro Kasai
Tokyo Electric Power Company Co., Ltd.
2-4-1 Nishi-tsutsujigaoka, chofu-shi, Tokyo, 182, Japan
ABSTRACT
A prototype of crack detection and meas-
urement system in which cracks are recog-
nized and the length, width and distribu-
tion of cracks are measured has been
developed for the purpose of assessment
of the deterioration of concrete struc-
tures. In the system images of different
resolution which are generated and hier-
archically structured from fine to coars-
er are processed with a single spatial
filter for crack detection rather than
using a series of spatial filters for an
image of single resolution.
Results show that 66 % of cracks of more
than 0.1 mm width have been detected with
width measurement accuracy of 0.08 mm
RMS.
KEY WORDS: Image processing, Crack
measuring, Edge detection, Spatial fil-
ter.
1.INTRODUCTION
To assess the deterioration of concrete
structures, it is important to inspect
cracks which can reveal useful informa-
tion for a diagnosis of a degree of
deterioration and also cause the other
deterioration such as carbonation, etc.
Recently a great deal of work has been
put in the inspection that a shape of
cracks is sketched out and a maximum
width of each cracks is manually measured
with crack-scale.
In order to save labor of the inspection,
image processing technique is applied to
recognition of a crack shape and to
measurement of a width. In this process-
ing the recognition and measurement
should be performed with accuracy of 0.1
mm RMS, as targeting the cracks ranging
from 0.1 to 3.0 mm in width. Hence "a
digital image examined in the process
should be so fine that the resolution of
image should be less than 0.025 mm per a
pixel.
In case of such a resolution an aspect of
3.0 mm width crack shows a band-like area
on the image, which is quite different
from a line-edge aspect of 0.1 mm width
crack. Detection of line-edge is usually
carried out by means of spatial filter-
ing. And the kernel size of filter is
selected in proportion to the width of
line-edge. For instance, a narrower
line-edge demands a smaller kernel, and a
wider like band area demands a larger.
Consequently, a series of spatial filters
must be prepared to detect a shape of
cracks with a variety of width. And a
larger kernel filter gives rise to ob-
struction of shortening a period of image
processing.
Hierarchical image processing algorithm
has been developed to detect cracks with
a variety of width by means of a single
spatial filter. The algorithm is divided
into two sub-algorithm. One is edge
detecting algorithm in which a series of
images with different resolution are
generated and structured hierarchically
from fine to coarser. And a single spa-
tial filter for line-edge detection works
on images in the hierarchy. The other is
crack measuring algorithm in which crack
shapes derived from images of different
resolution are combined and width of
every crack is calculated.
2.HIERARCHICAL EDGE DETECTING ALGORITHM
Edge detecting process is combined with
hierarchical image generating process in
which a series of images are structured
from fine to coarse in resolution.
2.1 GENERATION OF A SERIES OF IMAGES
A coarser image is generated from the
original fine image by means of 'reduc-
tion in size by selecting the maximum
value of pixels'. Fig. 1 shows a schema
of the reduction to 1/5 in size. The
reduced image is composed of the pixels
selected under the condition with the
maximum value remaining in each corre-
sponding 5x5 area on the original fine
image, because a pixel on crack has
higher value than on concrete.
22 | 24 | 35 | 45
24 | 25 | 45 | 45
n | 23 | 25 | 35 | 43
| 32 | 27 | 35 (47)
33 | 29 | 36 | 34
n
À fine image
A 1/5 reduced image
Fig. 1. Reduction in size by selecting maximum value of pixels
While cracks are reduced into 1/n scale
on coarse image, the aspect of band-like
cracks of n pixels in width generally
turns into line-edge. Although the cracks
of less than n pixels are reduced with
the same manner, the skeletonized cracks
are still remained by means of the maxi-
mum value selection. In the reduction,
the aspect of cracks less than n pixels
