Project Report Data
Project Report Data
Project Report Data
4, August 2013
representative morphologies of the sewer pipe defects from weighted median filter is similar to that of median filter but
the inspection images due to the noisy environments. To the pixel value at the central position of weighted median
effectively reduce the interferences of the noisy filter is duplicated k times. Thus, weighted median filter can
environments, this research designed the weighted median effectively remove environmental noise and remain texture
filters of different sizes to smooth the details of the noisy as completely as possible [10].
environments as well as remain the textures of the cracks.
B. Image Opening
An image opening is applied to the filtered images to
enhance the textures of the cracks. Considering the Image opening is one of the mathematical morphology
characteristic of irregular line-like structures of cracks, a operations to detect interested image region. Before an
structure element of line shape is given for the image opening, operation of image opening, the type of structural element
which is executed in the directions of 0°, 30°, 60°, 90°, 120°, needs to be determined. The determination of the type of
structural element depends upon the kind of interested shape.
and 150° for each gray level image. Overlapping the opening
In other words, the structural element is used to assist the
operated images of the different directions into a single one is
operation of image opening in enhancing the image regions
to be transformed into a binary image by Otsu’s thresholding.
similar to the structural element. Due to the morphological
After the binary transformation (or called image
characteristic of irregular line-like of crack, this research
segmentation), a large segmented region has high probability
introduces a linear structural element of 15 pixels into the
of being considered as a crack compared to a segmented
operation of image opening in the six directions, including 0°,
noisy with a relatively small region. In addition, the
30°, 60°, 90°, 120°, and 150°.
eccentricity values of 0 and 1 represent the shapes of circle
and line, respectively. A segmented region with an C. Otsu’s Thresholding
eccentricity approximating to 1 means high probability of Otsu's thresholding, which is a thresholding method based
being considered as a crack. Therefore, two morphological discriminant analysis, determines the optimal thresholds for
features, including area and eccentricity, are measured for image opening by maximizing the following measure of class
each segmented region. Based on the measured separability [11]:
morphological features, several criteria will be established
and used to determine either crack or non-crack pattern in P1 (T ) P2 (T )[m1 (T ) − m2 (T )]2
D(T ) = (1)
image. P1 (T )σ 12 (T ) + P2 (T )σ 22 (T )
458
IACSIT International Journal of Engineering and Technology, Vol. 5, No. 4, August 2013
examined the sensitivity of the wavelet techniques, including samples using the weighted median filters with k=1.
Gabor and Haar wavelets, in the detection of cracks in beam According to each criterion listed in the tables, the accuracies
structures [15]. The experimental result indicates that Haar of crack and non-crack detections were estimated. Under an
wavelets exhibit superior performance for detection of area threshold, the higher the given eccentricity threshold is,
discrete cracks. Naseralavi et al. (2011) also proposed a the lower the accuracy of crack detection is obtained. On the
method based on sensitivity analysis of the structure for contrary, the derived accuracy of non-crack detection is
structural damage detection using natural frequencies [16]. ascent once the higher eccentricity threshold is given. Thus,
Based on the above literatures, sensitivity analysis has been the crack and non-crack detections have a trade-off
demonstrated to be able to effectively offer optimal relationship.
parameters for decision problems. There are i criteria and j In Table II through Table VI, the weighted median filters
weights to be determined so to have i * j times of accuracy with the different weights k, including 1, 3, 5, 7, and 9, were
computation remaining for solving such problem. tested. The sensitivity analyses demonstrate that the weight k
for the weighted median filter has slight effect on crack
detection based on the insignificant difference among the
IV. EXPERIMENTAL RESULTS
best overall accuracies of 0.80 through 0.90. Moreover, six
In order to determine the optimal parameters, including criteria consisting of area (in pixel) threshold and eccentricity
weight of weighted median filter, area, and eccentricity, for threshold, including (≧50, ≧0.95), (≧100, ≧0.90), (≧100,
crack detection, a preliminary survey to the derived ≧0.95), (≧200, ≧0.80), (≧200, ≧0.85), and (≧200, ≧
morphological features is necessary. In the preliminary
0.90), can lead to the best overall accuracies under the
survey, the weight k was given as one, i.e. equivalent weight,
condition of various weights setting in filtering.
and applied to the two images in Fig. 1. The segmentation
results by the steps of A through C in section III are shown as TABLE I: NUMBER OF THE SEGMENTED REGIONS IN FIG. 3
Fig. 3. Compared with the two pairs of the segmentation, the Figure ID
larger weighted median filter significantly reduces the
Area (pixels) (a1) (a2) (b1) (b2)
number of the segmented regions of noisy environment.
≧200 5 4 0 0
Table I is a statistic list for the number of the segmented
200> ~ ≧150 1 0 0 0
regions in Fig. 3. Table I is seen that the segmented regions,
which consist of less than 50 pixels or have the eccentricities 150> ~ ≧100 3 2 0 0
of less than 0.8, are the majority for each binary image. 100> ~ ≧ 50 4 5 1 1
Unfortunately, most of those segmented regions belong to < 50 166 106 139 85
noisy environment. In other words, a segmented region Eccentricity
satisfying the criterion, i.e. area threshold ≧ 50 pixels 1.0≧ ~ ≧0.95 10 5 1 4
coupled with eccentricity threshold ≧ 0.8, has a higher 0.95> ~ ≧0.90 10 13 7 7
probability to be portion of crack. Based on the above result, 0.90> ~ ≧0.85 41 32 37 20
a sensitivity analysis was schemed for finding out the 0.85> ~ ≧0.80 6 2 3 1
appropriate criteria for crack detection. < 0.80 112 65 92 54
459
IACSIT International Journal of Engineering and Technology, Vol. 5, No. 4, August 2013
TABLE III: SENSITIVITY ANALYSIS FOR THE 50 TRAINING SAMPLES USING A threshold ≧ 0.90 are applied. The difference of 12%
5×5 WEIGHTED MEDIAN FILTER WITH K = 3
indicates that the testing samples compared to the training
Criterion Accuracy of Accuracy of Overall
ones have six omitted samples (three for crack detection and
crack non-crack accuracy
detection detection three for non-crack detection). On the contrary, the least
Area Eccentricity
(pixels)
difference of the overall accuracies between the training and
testing samples is 2% while k = 9, and area threshold ≧ 200
>=50 >=0.80 0.96 0.52 0.74
>=0.85 0.96 0.52 0.74
pixels coupled with eccentricity threshold ≧ 0.90 are
>=0.90 0.96 0.60 0.78
adopted. The overall accuracy difference of 2% is caused by
that the testing samples compared to the training samples
>=0.95 0.92 0.76 0.84
have three more omitted samples for crack detection but two
>=100 >=0.80 0.92 0.72 0.82
less omitted samples for non-crack detection. Consequently,
>=0.85 0.92 0.76 0.84
this paper demonstrates that the different criteria for crack
>=0.90 0.92 0.80 0.86
detection could result in slight omission errors that cause no
>=0.95 0.88 0.88 0.88 great effect on crack defection.
>=150 >=0.80 0.84 0.80 0.82
TABLE V: SENSITIVITY ANALYSIS FOR THE 50 TRAINING SAMPLES USING A
>=0.85 0.84 0.80 0.82 5×5 WEIGHTED MEDIAN FILTER WITH K = 7
>=0.90 0.84 0.84 0.84 Criterion Accuracy of Accuracy of Overall
>=0.95 0.76 0.88 0.82 Area Eccentricity crack non-crack accuracy
>=200 >=0.80 0.80 0.84 0.82 (pixels) detection detection
460
IACSIT International Journal of Engineering and Technology, Vol. 5, No. 4, August 2013
461