1st International Conference on 3D Construction Printing
Swinburne University of Technology, Melbourne, Australia
26th -28rd November 2018 (Monday – Wednesday)
Comparative Analysis of Five Edge Detection Algorithms for Panel
Defect Detection
Chang Liu, Sara Shirowzhan, Samad M. E. Sepasgozar, Seyedalireza Kaboli
Faculty of Built Environment, the University of New South Wales Sydney, Australia, 2052.
ABSTRACT
Identifying construction defects is an important task, which depends on construction
methods and quality control procedures. However, manual defect detection is time-
consuming especially when the external cladding should be controlled in terms of the
quality of panels and installations. Besides, the process of external cladding quality control
based on digital images and using vision-based algorithms have been received less
attention in the literature. This paper aims to present an algorithm able to detect panel
edges for controlling the quality and estimate the edges of panels during construction. Two
different buildings available to publics are chosen and images are taken for edge detection
analysis. Five proposed algorithms are chosen for processing the selected images and the
results are evaluated by employing two error and noise analysis methods Mean Square
Error and Peak Signal to Noise Ratio. The evaluation methods show that Canny algorithm
performs well and gives the best results among the five algorithms for panels’ edge
detection. The results show that Canny produces the lowest level of error and noise in the
selected panels. This study also shows that the accuracy of edge detection processes
depends on both the chosen algorithm and items contained in photos. The image
processing results show that the accuracy of algorithms could be improved by removing
unwanted items such as vegetation and sky clouds. The evaluated algorithms allow
practitioners to collect images in construction site and share them with their supervisors
who are controlling the quality of panel installation in a medium sized project. The future
work needs to focus on colour edge detection for automatic defect identification using RGB
and 360 images. In addition, the workflow for automatic detection and classification of defects
should be developed to be able to analyse different panels in terms of textures, size and
installation mechanisms.
http://3dcpconference.com/
1st International Conference on 3D Construction Printing 2
1. Introduction
With the development of digital technology, automatic procedures are widely used in
various construction applications. While automatic defect detection are necessary in
construction, defect detection is usually undertaken in a manual operation. This is a
time-consuming procedure which increases construction cost. Besides, with the
increasing use of digital data including images and point clouds recently in construction
projects (Hamledari et al., 2017, Shirowzhan et al., 2018, Sepasgozar et al., 2018), edge
detection has been a popular field of study for defect detection in engineering. For
example, it has been used in robotic applications that helps to replace human eyes for
measurement and judgement (Jain et al., 1995). It also has been used for handwriting
and face recognition for a long time (Bhowmik et al., 2017, Yang and Huang, 1994). It
is also gaining more interest in civil engineering studies and applications. For instance,
there is an increasing trend towards applications in earthquake damage estimation of
buildings and defect detection in infrastructure. One example is that Huang et al. (2012)
applied edge detection for building configuration measurement in seismic damage
estimation. Another example is that Tahmid and Hossain (2017) applied vehicle edge
detection for smart traffic control. Thus, automatic edge detection could be a time saving
approach for defect detection in external cladding. Scratches to the face of panel, panels
sitting in or out, and having unequal joints in an elevation are the main defects in external
cladding. Huang et al. (2014) used edge detection algorithms for pavement crack
detection. They collected 2D grey-scale images combining with 3D laser scanning
information with 2D cameras and laser line projectors mounted in a car. As surroundings
are lighter than pavement images in grey-scale images, pavements can be detected by
judging brightness. Laser information contains depth information that is used for finding
defects after knowing where pavements are. Cha et al. (2017) used convolutional neural
networks (CNNs) and deep learning for concrete defect detection and crack damage
detection based on edge detection. 332 raw concrete images have been selected and
the results of their image analysis were highly accurate by detecting 90% of cases. Chen
et al. (2017) proposed a video processing method with Bayesian data to detect crack
edges on metallic surfaces of nuclear power plants. They suggested that their
algorithms in C++ showed better performance than MATLAB. Feng et al. (2017)
developed a deep active learning system for automatic concrete defect detection
including cracks, deposits and water leakages. They achieved 87.5% accuracy.
Although there have been a few civil infrastructure studies on defect detection of
concrete and asphalt pipe change monitoring such as the one by (Koch et al., 2015),
research on panel defect detection is rare in construction. Thus, the current gap in
literature is the performance of the current algorithms for different types of digital images
including 3D and 360s can only be used in certain situations in terms of their
performance for quality control in different construction contexts such as modular
construction. In other words, defect detection methods vary for different items and thus
defect detection methods of other objects such as concrete pipes cannot be used for
panels. Hence, the aim of this paper is finding an edge detection algorithm for panel
defect detection comparative analysis of five edge detection algorithms.
Three-dimensional (3D) camera is usually used for data collection as it can provide more
information than ordinary two-dimensional (2D) camera including panorama and depth
3 Comparative Analysis of Five Edge Detection Algorithms for Panel Defect Detection, Liu et. Al.
of image. For example, in (Sun et al., 2017), 3D camera is used for room views to judge
room price. Thus, it is an ideal technique for 3D model data collection. In this paper, 3D
camera is adopted, but only used for obtaining RGB pictures. Other information in it is
for further research by authors in future. The literature shows that Roberts, Prewitt,
Sobel, Laplacian of Gaussian (LoG) and Canny are the most common algorithms in
different contexts. This study adopts these methods and will evaluate their performance
in analysing panel installation quality. Mean Square Error (MSE) and Peak Signal to
Noise Ratio (PSNR) are two ratios to judge accuracy of these algorithms. This paper
presents the research methodology including analysis of MSE and PSNR of five
algorithms and discussion of omission and commission errors in algorithms. Field
experimentation contains two construction sites with eight chosen places. Results and
conclusion show that different algorithms can be used in different situations.
2. Research Methodology
In order to analyse the quality of panels, a 3D camera was used to collect images from
panels of two selected buildings. Five key algorithms were employed to process the
images and detect edges of panels. The results of the applied algorithms were also
analysed using an error analysis method and a noise ration analysis method. The
algorithms and performance analysis methods are discussed in this section. Edge
detection algorithms produce edge points or edge fragments from a picture (Jain et al.,
1995). Table 1 shows the five edge detection algorithms chosen for this study Sobel,
Canny, LoG, Prewitt and Roberts, which were operated in MATLAB. As for detailed
calculation steps of these edge detection algorithms, filtering and calculating gradient
magnitude are two main parts. Filtering is a main part is because that weighted average
of images is necessary for calculating gradients of edges of grey-level edge detection
and it is calculated with the filter. The reason of importance of calculating gradient
magnitude is that edges are found by gradient magnitude of calculated weighted
average of images (Jain et al., 1995). To be specific, filtering can enhance accuracy of
calculating gradients of edges and decrease noises in image. After filtering, edges can
be detected by calculating the gradient magnitude of two adjacent points with designed
threshold. Another point needs to mention is that every filter used for edge detection in
convolution conclude two parts, vertical ones and horizontal ones. In order to compare
the performance of the proposed different edge detection operators, both MSE and
PSNR methods were employed. The lower MSE and higher PSNR is better for
performance. For example, Acharjya et al. (2012) used MSE and PSNR for comparing
the algorithms performances of face images.
Table 1. Utilised edge detection filters.
Method Algorithm Descriptions
Roberts (Roberts, “v” and “h” mean vertical
G=�vRoberts 2 +hRoberts
2
1963) and horizontal parts of
the filter. They are 2*2
matrixes.
Prewitt (Prewitt, The filter consists of 3*3
G=�vPrewitt 2 +hPrewitt
2
1970) matrixes.
1st International Conference on 3D Construction Printing 4
Sobel (Sobel and Emphasis on pixels that
G=�vSobel 2 +hSobel
2
Feldman, 1968) near to the centre of the
3*3 matrixes.
2 2
LoG (Marr and 1 x2 +y2 -x +y2 As the filter is
Hildreth, 1980) LoG(x,y)=- 4 �1- � e 2σ symmetric, vertical and
πσ 2σ2
horizontal parts are the
Combination of Gaussian and
same.
Laplacian filers.
Canny (1986) Smoothing images by a certain filter, Thresholds of Canny are
finding intensity gradients, applying set in MATLAB
non-maximum suppression, automatically.
choosing minimum and maximum
and connecting edges.
Note: G refers to Gradient magnitude.
The equations for MSE and PSNR are provided below based on (Lehmann and
Casella, 2006) respectively.
m n
1 2
MSE= � � (Iprocessed (i,j)-Igrey (i,j))
m×n
i=1 j=1
2552
PSNR=10×log10 � �
MSE
Where m and n are the length and width of the image, Igrey (i,j) represents value of pixel
(i,j) in original image, and Iprocessed (i,j) represents value of pixel (i,j) in processed image.
3. Field Experimentation and Results
Two buildings were selected for analysing their panel installation quality. total number
of images for all eight chosen study areas is 100.
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 1. Eight different panels chosen for evaluating the image processing methods.
Note: Panels were chosen from different parts of selected buildings including southern
side, ground floor and western corner.
Figure 1 shows different types of panels used in different locations, indicated by a to h.
For example, Figure 1 (b) contains arch panels for judging whether the curve edges are
properly installed in the correct location. Figure 2 shows the results of edge detection
analysis for the Figure 1 (a) using the five chosen algorithms. Figure 2 (a, b and c) show
the main edges of panel and Figure 2 (e) shows more unwanted edges especially left
corner but fewer discontinuous lines than Figure 2 (a). Besides, human eyes may
5 Comparative Analysis of Five Edge Detection Algorithms for Panel Defect Detection, Liu et. Al.
consider edges in the left corner as noises, but in fact they are known as real edges in
the computer language. Figure 2 (e) detect most edges from original figure and panel
edges can been shown clearly, but the same problem as Figure 2 (d) still exists such as
presenting too many edges including tree edges. All these five algorithms have the
problems that no edge places are shown as edges or edges are not be shown. To
measure difference in the performance of the algorithms, MSE and PSNR are calculated
for difference between processed and original images and noise level respectively.
Based on the average value calculated for the sample of images collected, MSE of
Canny is lowest and PSNR of Canny is highest. MSE of Roberts is highest and PSNR
of Roberts is lowest. The lower MSE and higher PSNR, the more accurate of the
algorithm. Thus, it could be judged that results of Canny performance are best and
Roberts performances can be considered as worst in this context. The results also show
that the best performing algorithms based on MSE are confirmed by the values achieved
for PSNR, which are Canny followed by LOG.
(a) Roberts (b) Prewitt (c) Sobel (d) LoG (e) Canny
Figure 2. Edge detection algorithm results of Figure 1 (a).
4. Discussion
The contribution of this paper lies in the implementation of five different algorithms on
panel installation quality control. Five different image processing algorithms were
applied to analyse our samples of selected panels from two different buildings. In order
to compare the performance of the chosen algorithms two main error measurement
methods MSE and PSNR were selected and applied. The result of MSE analysis shows
that Canny has the lowset value in all samples. This is in converse for PSNR. This shows
that this algorithm has higher accuracy in analysing images from panels. However,
Canny shows more useless edges than others because in this paper only panel edges
are expected to be detected. All these five algorithms can only detect edges but no
classification of edges. Hence, it is better to choose other algorithms that can distinguish
panel edges and other edges in future research. Thus, it is hard to show the results for
both Sobel and Roberts are the same and cannot be distinguished. The results also
show that unwanted items such as vegetation and cloudy sky produce a lot of curved
lines, so they may decrease the accuracy of an edge detection algorithm. This implies
that the method of photography (e.g. view angle) affects the photo quality for including
unwanted items.
If only simple edges such as straight lines exist in one figure, accuracy will be higher
compared to the images including curved, straight and angled lines and vegetation.
Figure 3 (a) presents an example of lower accuracy edge detection, since trees as
unwanted items will affect the image analysis. In contrast, almost all lines in Figure 3 (b)
are straight and there are no unwanted items beside panels. MSE and PSNR of Figure
3 (b) have much better results than Figure 3 (a). To be specific, Figures 3 (a) and (b)
contain different materials such as aluminium sheet, aluminium composite and CFC
1st International Conference on 3D Construction Printing 6
panels which could be used as external cladding. The length and width of raw sheet
vary between 1m to 1.5m, and 2.5m to 4m respectively. Figure 3 (c) shows another
example of unwanted items (i.e. cloudy sky) in an image that makes lower accuracy
results. By contrast, Figure 3 (d) includes a photo with lower unwanted items that already
achieved higher rate of accuracy. Another important point for discussion could be view
angle in photography that helps exclude unwanted items such as the way that Figure 3
(d) was taken. It was found that the amount of clouds may affect the brightness of an
image that leads to changes of the accuracy of the algorithms. We recommend
performing a future study to confirm this effect. Moreover, analysis results show that
edge detection algorithms can be used for panel detection and Canny operator shows
the best performance for panel construction, because Canny operator shows best
performance in every photo evaluated by using MSE and PSNR methods.
MSE: 100.08 MSE: 71.28 MSE: 161.08 MSE: 97.69
PSNR: 28.16 PSNR: 29.64 PSNR: 26.09 PSNR: 28.27
(a) (b) (c) (d)
Figure 3. Comparison of Canny algorithm accuracy of four photos.
The practical implication of this study is that less skilled labours can collect digital
images from construction sites when installing panels and send it to the head office for
analysing the quality control remotely. The presented methods can assist to identify the
shapes of defects including abrupt changes in small areas to the straight lines expected
for panels. There is also some recommendation about the cladding zone and joint sizes
which should be followed during the panel installation. This is important to have a plum
sub-framing, otherwise the panels might be installed in or out. Joint sizes might also be
different between two projects depending the architect’s comments and vary between
10mm to 15mm. One of the mistakes which might happen during the panel installation
is having different joint sizes from what given in the drawings. To avoid unequal joint
sizes in the cladding, the panels are required to be installed in the correct location. The
later we find the mistakes on site, the costlier they could be because we are required to
rectify the issue for all those installed panels, especially if all the panels are already
fabricated based on the drawings. This would be even worse if we find the issue after
dismantling the scaffold because the rectification should be done through boom/scissor
lift which is time consuming. To find those issues in the early stages, we are required to
visit site every day and spend plenty of time on each project which increases the
construction costs. However, using a camera by a less skilled person, all panels can be
captured and an automatic edge detection system may assist in finding the defects. This
approach may reduce the time a skilled technician needs to spend on site and the
construction costs consequently.
7 Comparative Analysis of Five Edge Detection Algorithms for Panel Defect Detection, Liu et. Al.
5. Conclusions
After comparative analysis of five edge detection algorithms, this paper states that
Canny would be a good choice for panel edge detection in external cladding during
construction. Thus, defects can be detected by observing results of edge detection in
office without going to construction site. Using this method, the edges can be detected
easily, and the contractors and engineers may find if the panels are installed according
the shop-drawings and there are equal joint sizes through the cladding. They will also
find if the panels are installed in place or they are sitting in or out. A 3D camera is used
for data collection with choosing eight places of two construction sites as case studies,
which is different from ordinary edge detection experiments using 2D camera. Besides,
MSE and PSNR results show that the accuracy of algorithms not only depend on
algorithms themselves but also items in photo. However, it still cannot achieve automatic
defect detection at this stage. Thus, future study may focus on improving algorithms and
proposing an automatic defect detection method. The future work needs to focus on
colour edge detection for automatic defect identification using RGB and 360 images.
Another part of the future study would focus on automatic categorization of different
types of panel defects.
References
Acharjya, P P, Das, R & Ghoshal, D 2012, 'Study and comparison of different edge
detectors for image segmentation', Global Journal of Computer Science and
Technology.
Bhowmik, S, Sen, S, Hori, N, Sarkar, R & Nasipuri, M 2017, 'Handwritten devanagari
numerals recognition using grid based hausdroff distance', Computer,
Communication and Electrical Technology: Proceedings of the International
Conference on Advancement of Computer Communication and Electrical
Technology (ACCET 2016), West Bengal, India, 21-22 October 2016. CRC
Press, pp. 15.
Canny, J 1986, 'A computational approach to edge detection', IEEE Transactions on
Pattern Analysis and Machine Intelligence, no.6, pp.679-698.
Cha, Y J, Choi, W & Büyüköztürk, O 2017, 'Deep learning‐based crack damage
detection using convolutional neural networks', Computer‐Aided Civil and
Infrastructure Engineering, vol.32, no.5, pp.361-378.
Chen, F C, Jahanshahi, M R, Wu, R T & Joffe, C 2017, 'A texture‐based video
processing methodology using bayesian data fusion for autonomous crack
detection on metallic surfaces', Computer‐Aided Civil and Infrastructure
Engineering, vol.32, no.4, pp.271-287.
Feng, C, Liu, M-Y, Kao, C-C & Lee, T-Y 2017, 'Deep active learning for civil
infrastructure defect detection and classification', Computing in civil engineering
2017.
Hamledari, H, Mccabe, B & Davari, S 2017, 'Automated computer vision-based
detection of components of under-construction indoor partitions', Automation in
Construction, vol.74, pp.78-94.
Huang, J, Liu, W & Sun, X 2014, 'A pavement crack detection method combining 2d
with 3d information based on dempster‐shafer theory', Computer‐Aided Civil
and Infrastructure Engineering, vol.29, no.4, pp.299-313.
1st International Conference on 3D Construction Printing 8
Huang, L, Oguni, K & Hori, M 2012, 'Image analysis of measuring building
configuration for seismic damage estimation', Natural hazards review, vol.14,
no.1, pp.1-10.
Jain, R, Kasturi, R & Schunck, B G 1995, 'Machine vision', McGraw-Hill New York.
Koch, C, Georgieva, K, Kasireddy, V, Akinci, B & Fieguth, P 2015, 'A review on
computer vision based defect detection and condition assessment of concrete
and asphalt civil infrastructure', Advanced Engineering Informatics, vol.29,
no.2, pp.196-210.
Lehmann, E L & Casella, G 2006, 'Theory of point estimation', Springer Science &
Business Media.
Marr, D & Hildreth, E 1980, 'Theory of edge detection', Proc. R. Soc. Lond. B, vol.207,
no.1167, pp.187-217.
Prewitt, J M 1970, 'Object enhancement and extraction', Picture processing and
Psychopictorics, vol.10, no.1, pp.15-19.
Roberts, L G 1963, 'Machine perception of three-dimensional solids', Massachusetts
Institute of Technology.
Sepasgozar, S M, Forsythe, P & Shirowzhan, S 2018, 'Evaluation of terrestrial and
mobile scanner technologies for part-built information modeling', Journal of
Construction Engineering and Management, vol.144, no.12, pp.04018110.
Shirowzhan, S, Sepasgozar, S & Liu, C 2018, 'Monitoring physical progress of indoor
buildings using mobile and terrestrial point clouds', Construction Research
Congress 2018.
Sobel, I & Feldman, J 1968, 'A 3 × 3 isotropic gradient operator for image processing',
Presented at a talk at the Stanford Artificial Project.
Sun, T, Xu, Z, Yuan, J, Liu, C & Ren, A 2017, 'Virtual experiencing and pricing of room
views based on bim and oblique photogrammetry', Procedia Engineering,
vol.196, pp.1122-1129, https://doi.org/10.1016/j.proeng.2017.08.071.
Tahmid, T & Hossain, E 2017, 'Density based smart traffic control system using canny
edge detection algorithm for congregating traffic information', Electrical
Information and Communication Technology (EICT), 2017 3rd International
Conference on. IEEE, pp. 1-5.
Yang, G & Huang, T S 1994, 'Human face detection in a complex background', Pattern
recognition, vol.27, no.1, pp.53-63.
