Early Defect Detection in Conveyor Belts using Machine Vision

Guilherme G. Netto

1,3 a

, Bruno N. Coelho

1,3 b

, Saul E. Delabrida

2 c

, Amilton Sinatora

3 d

ector Azp

urua

3 e

, Gustavo Pessin

3 f

, Ricardo A. R. Oliveira

2 g

and Andrea G. C. Bianchi

2 h

Department of Engineering of Control and Automation, School of Mines, Federal University of Ouro Preto (UFOP),

122 Diogo de Vasconcelos, Ouro Preto, MG, 35400-000, Brazil

Computing Department, Federal University of Ouro Preto (UFOP),

122 Diogo de Vasconcelos, Ouro Preto, MG, 35400-000, Brazil

Vale Institute of Technology (ITV), 31 Juscelino Kubitschek, Ouro Preto, MG, 35400-000, Brazil

Keywords:

Image and Signal Processing, Curvature Outlier, Defects Detection, Machine Vision Inspection, Maintenance.

Abstract:

Continuous belt monitoring is of utmost importance since wears on its surface can develop into tears and

even rupture. It can causes the interruption of the conveyor, and consequently, loss of capital, or even worse,

serious or fatal accidents. This paper proposes a laser-based machine vision method for detecting defects in

conveyor belts to solve the monitoring problem. The approach transforms an image of a laser line into a one-

dimensional signal, then analyzes it to detect defects, considering that variations in this signal are caused by

defects/imperfections on the belt surface. Differently from previous works, the proposed method can identify

a defect through a 2D reconstruction of it. The results reveal that the proposed method was capable to detect

superﬁcial imperfections in simulated conveyor belt experiments, achieving high values in metrics such as

precision and recall.

1 INTRODUCTION

Conveyor belt has been developed and used for

decades as an essential part of transportation pro-

cesses (Pang and Lodewijks, 2005), it can be cited

as the most cost-effective equipment for the contin-

uous transport of large amounts of material due to

their high efﬁciency, large capacity, relatively simple

construction, and less maintenance required (Fedorko

et al., 2014; fu Hou and rui Meng, 2008). A problem

found in traditional monitoring systems for conveyor

belt is operator dependence. At the same time, rollers

and pulleys can only be monitored while they are run-

ning. On the other hand, conveyor belts can only be

observed when they are not in operation. Also, it is

often difﬁcult to deﬁne whether a malfunction was

caused by the conveyor’s operation or by the act of

https://orcid.org/0000-0001-7255-5708

https://orcid.org/0000-0002-2809-7778

https://orcid.org/0000-0002-8961-5313

https://orcid.org/0000-0002-5004-4947

https://orcid.org/0000-0002-4182-2540

https://orcid.org/0000-0002-7411-9229

https://orcid.org/0000-0001-5167-1523

https://orcid.org/0000-0001-7949-1188

taking the sample when there is a need for a destruc-

tive test on the belt (Fedorko et al., 2018).

Conveyor belts are prone to failures, such as: wear

on the surface, degradation due to weathering, dam-

age to the ends and longitudinal and perpendicular

tears in sections of the belt, misalignment of the belt

during operation, overheating of the rollers, among

others, causing production risk (Yang et al., 2014).

Failures in such equipment systems generally result

in a production stoppage for repair, or in more serious

cases, as accidents with workers, resulting in serious

consequences to the company.

Traditional defect techniques and instruments usu-

ally require physical contact with parts of the con-

veyor to identify faults, or worst, just identify faults

only when they occur, like a tear that crosses the

thickness of the belt, causing material leaking. An ex-

ample is a mechanical sensor as a misalignment sen-

sors, which are composed of position switches that

trigger an alarm when in contact with the misaligned

belt. Depending on the severity of the misalignment,

stop the conveyor operation.

There are some sophisticated techniques for mon-

itoring conveyor belt components such as systems

based in RFID sensors (Pang and Lodewijks, 2006),

magnetic sensors (Nicolay et al., 2004), X-rays sen-

Netto, G., Coelho, B., Delabrida, S., Sinatora, A., Azpúrua, H., Pessin, G., Oliveira, R. and Bianchi, A.

Early Defect Detection in Conveyor Belts using Machine Vision.

DOI: 10.5220/0010240803030310

In Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2021) - Volume 4: VISAPP, pages

303-310

ISBN: 978-989-758-488-6

303

sors (Yang et al., 2016) among others. It is notable

that these monitoring methods, even though some

have a better monitoring capacity than standard meth-

ods, need a special construction, with sensors inside

the belt, or unusual equipment used in industrial ar-

eas, such as an X-ray emitter. Thus, image analysis

using computer vision has been a strong candidate for

this kind of approach.

This paper aims to present an automatic method

based on computer vision, clustering and statistical

basic techniques capable of detecting and describe

defects in the conveyor belt, as well as rebuild a

2D shape of those identiﬁed defects. The proposed

method can identify tears, bubbles, and wear on the

belt surface. Instead of using regular laser belt sur-

face scanning, this approach uses curvature outlier

values of a one-dimensional signal extracted from a

laser line’s image to represent defects through the belt

surface. The major contributions of this work are as

follows: a new machine vision approach to detect a

defect in conveyor belts using a camera and a laser,

interpretability with a 2D shape of those identiﬁed de-

fects and recognition of various types of defects, such

as bubble, wear and rips.

This paper is organized as follows: Section 2 dis-

cusses the techniques used in the literature to deal

with the same problem. Section 3 introduces the pro-

posed materials and method. Section 4 presents a

short explanation of the database and method evalua-

tion. Section 5 presents the experiments and obtained

results. Then, the conclusions and future works are

exposed in Section 6.

2 RELATED WORKS

Sophisticated and intelligent techniques that monitor

the belt conveyor’s status have already been devel-

oped. (Nicolay et al., 2004) proposed a system that

made use of RFID using tags spread over the entire

length of the belt, making it or a very long time large

between readings or the lack of a tag would trigger the

tearing alarm on the belt. (Pang and Lodewijks, 2006)

proposed using magnets inside the belt and using an

external sensor, making it determine the belt condi-

tions such as: operating speed, wear, among others.

(Guan et al., 2008) proposed the use of X-rays in con-

junction with a receiver so that by observing the atten-

uation of the signal received by the receiver, the sys-

tem was able to identify ruptures in the steel webs of

the belts. Considering the overheating, (Nascimento

et al., 2017) proposed using computer vision in an

unmanned aerial vehicle (UAV) for the inspection of

rollers through the image of a thermal camera.

(Fromme et al., 2006) proposed a system capable

of detecting defects in sections of the belt using data

from at least one camera, using a encoder to adjust the

camera frames. (Kurihara et al., 2006) developed an

apparatus using a camera and laser to detect tears in

conveyor belts by spacing the laser beam in the case

of a tear in the belt.

(Li et al., 2011) proposed an intelligent method

using characteristics of the binary image of the belt

(such as shape, position and size of the identiﬁed

objects) together with ANDs and ORs to determine

whether or not the belt has tears. (Peng, 2013) pro-

posed using the MDNMS algorithm to identify tears

in grayscale images of conveyor belts and developing

a method of automatic false positive detection using

the classiﬁer AdaBoost.

(Yang et al., 2014) developed an apparatus using

a camera focused on a brightly lit area, providing a

system capable of identifying tears and misalignment

in conveyor belts. (Li and Miao, 2016) proposed us-

ing the SSR algorithm to identify slots present in con-

veyor belts through characteristics of the resulting bi-

nary image, such as boundary areas of rectangles of

the identiﬁed objects, among others

(Yang et al., 2016) proposed to transform the two-

dimensional signal obtained by the camera into a one-

dimensional vector, and from there, detect the tears

in conveyor belts through the difference between the

intensities obtained by the one-dimensional vector.

(Qiao et al., 2016) proposed using two cameras (in-

frared and CCD) and a laser beam emitted on the belt,

detecting tears through the discontinuity of the laser

by the CCD, and the histogram obtained by an in-

frared camera. (Qiao et al., 2017) used a camera in

conjunction with a laser beam emitted on the belt sur-

face to detect tears in conveyor belts, using edge de-

tection and Hough transform.

Although the aforementioned works can identify

defects such as longitudinal tears in conveyor belts,

there are still few, if any, that can inform any other

information or details about other types of defects,

such as surface erosion, for example. To inform more

about different types of defects detected, such as bub-

bles and surface erosion, the present work proposes

to use a camera focused on a section of the belt where

a laser beam is used perpendicular to the direction of

movement of the belt, assuming that defects present in

the belt proﬁle cause deformations in the shape of the

laser captured by the camera, and, capturing these de-

formations, identify and reconstruct the shape of these

defects.

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

304

3 MATERIALS AND METHODS

To identify defects covering belt surfaces, the exper-

imental set up used a camera and a laser line out

of phase and positioned below the lower belt. A

schematic representation is shown in Figure 1. The

hypothesis assumes that with this conﬁguration, a

laser line on the belt surface can emphasize irregu-

larities present on it, even if not visible to the naked

eye. When combined with automatic computer vision

techniques, it can automatically detect defects.

Figure 1: Schematic representation of the experimental

setup for image acquisition.

The ﬁrst step is image acquisition and laser seg-

mentation as an object of interest. Then, the object is

transformed into a one-dimensional signal, and its 1D

curvature is calculated. The statistical analysis iden-

tiﬁes outlier points of curvature, assuming that such

points are caused by irregularities on the belt surface,

and consequently in the one-dimensional signal. The

discrepant points are grouped using an unsupervised

algorithm, and a bounding box is drawn around them,

a 2D defect representation, the identiﬁcation itself.

The following sections present the description of the

proposed method in detail.

3.1 Laser Segmentation

In this work, the more we save time processing, the

better. So, we proposed to simulate a system that

could use a monochromatic digital camera with a

specter ﬁlter. Since we proposed to use a laser, the

ﬁlter must be at the same specter frequency of the

chosen laser. This way, we could achieve the same

result of an RGB-Gray color transformation without

any computational cost. The result is shown in Fig-

ure 2a, a grayscale image with high intensity was the

laser is present and low intensity in the remainder of

the image.

Bilateral Filter (Elad, 2002), a non-linear ﬁlter ca-

pable of reducing noises without loss of image infor-

mation was used. The ﬁltering result is a smooth im-

age without most of the acquired noises, keeping the

format of the laser. The bilateral ﬁlter has two pa-

rameters: spatial (σ) and range (r). The spatial pa-

rameter is proportional to the size of the image, and

the range parameter is proportional to the amplitude

of the edges. After the non-linear smoothing, Otsu’s

Method (Otsu, 1979) was used to binarize the image,

since it has an excellent performance in bimodal im-

ages. The result is an image with 0s representing fore-

ground and 1s for laser line, as shown in Figure 2b.

(a) R-channel of the RGB image

(b) Resulting binary image after preprocessing

Figure 2: Gray and preprocessed image, respectively.

3.2 One-dimensional Signal

Transformation

Due to the laser’s width, the segmented line has a

width as well. Furthermore, as an image is usually

toughest to process than a one-dimensional signal, we

propose to transform it in a one-dimensional signal

named y.

Consider the image denoted by I(i, j), of size M ×

N, where i and j represents the rows and columns of

the image, respectively, being i = 1 to M, and j =

1 to N. The one-dimensional transformation occurs

assigning to each column j of the image I the mean

value of all the indices i of each row when I(i, j) =

1. Algorithm 3.1 presents the pseudo-code for one-

dimensional transformation y( j) with j varying from

1 to N, i.e., image width now representing the length

of transformed signal. Figure 3 shows the result of the

one-dimensional transformation of the binary image

of Figure 2b in a signal y with size N.

3.3 Curvature Calculation

It is well accepted that curvature provides an essen-

tial representation of salient shape points, as well as

Early Defect Detection in Conveyor Belts using Machine Vision

305

Algorithm 3.1: One-dimensional transformation.

1 for j = 1 → N do

2 m ← 0; cont ← 0;

3 for i = 1 → M do

4 if I(i, j) 6= 0 then

5 m ← m + i;

6 cont ← cont + 1;

7 end

8 end

9 if cont = 0 then

10 cont ← 1;

11 end

12 y( j) =

cont

;

13 end

Figure 3: One-dimensional signal.

invariant to rigid-body transformations (Estrozi et al.,

2003). Once the one-dimensional signal y was ob-

tained, a numerical curvature κ of it was calculated

through Equation 1

κ =

(1 + y

)

(1)

where y

is the ﬁrst and y

the second derivatives of

one-dimensional signal. The complete description of

numeric curvature calculation is explained in (Xian-

guo et al., 2018) and (Junior and Costa, 1996).

Figure 4 presents the curvatures κ as function of

length of one-dimensional signal y for two different

image frames (belt position), where the blue curve

represents the calculated curvature in a region with

defect and red without one.

Figure 4: Curvature of the section with (blue) and without

(orange) defect.

As shown in Figure 4, both curvatures are very

similar, except in the location of the defect, region of

the curvature peak in the blue curve. Consequently,

the main hypothesis is that outlier curvature values

may represent defects at the belt surface.

A closer look at Figure 4 shows that in the regions

close to the defect, the curvature values are higher, or

an outlier in that exact frame, whereas in the remain-

ing part of the signal, the curvature is relatively simi-

lar. We did not use a static threshold value because it

is difﬁcult to choose a curvature value for a belt sur-

face representing a defect on it since the same belt

may have different surfaces: straight in a location and

noisier in another one. Therefore, a curvature point

representing a defect in the straight section of the belt

may not represent in the noisily. Thus we propose

to use outlier curvature points as a descriptor of the

defects. In other words, the outlier will represent a

curvature value that diverges from an overall pattern

on a sample.

In data processing, there are many techniques for

outlier detection. In this work, four known statistical

techniques were used: Z-Score (Hodge and Austin,

2004), Chauvenet Criterion (Lin and Sherman, 2007),

Median Absolute Deviation (MAD) (Leys et al.,

2013) and Interquartile Range (IQR) (Jeong et al.,

2017). The ﬁrst two use concepts of mean and stan-

dard deviation, whereas the last two use quantile con-

cepts. As every technique has particularities, to use

all of them was proposed, and accept an outlier detec-

tion only if three of them detect the same point. The

detection becomes more robust to noises, as at least

one of the two concepts (mean+standard deviation or

quantile) can detect the same outliers. Moreover, in

data processing, usually, the outliers are removed or

replaced in the data. However, in this work, we iden-

tify these outliers’ index and assign these points as

defects.

3.4 Defect Detection

Many frames with defect information in the laser line

will be captured depending on the velocity of the con-

veyor and the size and shape of the defect, so many

outlier curvature points can be recognized. These

points are allocated into a defect matrix, with the

same dimensions rows×columns of the image, in the

correspondent column j that they were found in the

signal, but updating the row value in every frame ac-

cording to the velocity of the conveyor. After the laser

passes through a complete defect, the points identiﬁed

has an approximate shape of the defect, which was de-

ﬁned as the 2D reconstruction of the defect.

The Density-Based Spatial Clustering of Applica-

tions with Noise (DBSCAN) algorithm (Ester et al.,

1996), which has the capacity of clustering data ac-

cording to the density of points in a spatial location,

makes the defect detection itself. The algorithm ana-

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

306

lyzes the defect matrix, grouping the nearby points as

a single defect. As DBSCAN is an unsupervised ma-

chine learning clustering algorithm, it does not need

previous training. Also, as it groups points locally, it

does not need to know the number of clusters inside

the data. Therefore, it makes the defects detection

possible, even not knowing the shapes or number of

defects present in the belt surface.

4 DEFECTS DATABASE AND

METHOD EVALUATION

The core of this work is to propose a method that

even with low-cost techniques, should be possible to

achieve a good result detecting defects in conveyor

belts, so before the implementation of the solution on

real conveyor belts, we tested our hypotheses on sim-

ulated images. The images samples were illuminated

by laser, according to the scheme presented in Fig-

ure 1. The database were constructed to simulates the

movement of a conveyor belt with defects illuminated

by a laser and ﬁlmed by a camera, where videos sim-

ulating the laser, and the movement of the conveyor

were recorded.

4.1 Synthetic Database

The synthetic database was developed using 3D

Studio Max



software. The environment was

simulated using a virtual conveyor belt with

4500mm×900mm×20mm. A video simulating the

movement of the conveyor was built with 30 seconds

of duration and 30 frames per second. To investigate

the conﬁgurations between camera and laser, many

simulations with the same video characteristics were

built, as duration and frame rate, but changing the po-

sitions of the laser and the camera. The laser was

lagged in 30

◦

, 45

◦

, and 60

◦

degrees of the camera,

whereas the camera was positioned perpendicular to

the belt. Also, the reverse of these conﬁgurations was

used as well.

To simulate defects, some irregularities (visible

and invisible with the naked eye) of many shapes and

sizes were inserted randomly in the simulated belt im-

age. The regions with the irregularities are shown in

Figure 5. These irregularities intend to describe real

defects such as rips, erosion, tears, wears, among oth-

ers.

4.2 Performance Evaluation

Considering the proposed method for defect detec-

tion, its evaluation is based on three known pattern

Figure 5: Examples of different irregularities proposed in

the simulation.

recognition metrics: Precision, Recall, and F1 Score.

These metrics measure how good the method’s detec-

tion was compared with the ground truth of the syn-

thetic database. Precision and recall are metrics use-

ful for measure relevance (Perruchet and Peereman,

2004), since they show the amount of data obtained

that are relevant (precision), and the amount of rel-

evant data obtained (recall). The Equations 2 and 3

describe precision and recall, respectively

Precision =

T P

T P + FP

(2)

Recall =

T P

T P + FN

(3)

where TP is the true positive, corresponding to defects

that were correctly detected, FP, the false positive, a

defect that was detected incorrectly, and FN, the false

negative, indicating no detection when it does.

Using precision and recall, we can calculate the F1

Score (Sasaki et al., 2007), which represents the bal-

ance between the two metrics, and can be described

by Equation 4.

F1 = 2 ∗

Precision ∗ Recall

Precision + Recall

(4)

5 EXPERIMENTS AND RESULTS

The proposed method was developed in Python pro-

gramming language, running Ubuntu 16.04 64 bits as

operational system. All experiments were performed

on a DELL Inspiron 15 7559, Intel



Core

i7 −

6700HQ processor with a 2.60Hz × 8, SSD 120Gb,

16Gb RAM.

Tests to obtain the best conﬁguration between

laser and camera were performed comparing the re-

sults obtained by the method with the ground truth.

Table 1 presents the mean values of precision, recall,

and F1 achieved by every conﬁguration between cam-

era and laser. The best conﬁguration was obtained

with the camera and laser lagged by an angle of 30

◦

(camera perpendicular to the belt).

Using the best conﬁguration, the experiment used

the synthetic database presented in Section 4 to obtain

Early Defect Detection in Conveyor Belts using Machine Vision

307

Table 1: Mean results of precision, recall and F1 for defects

detection in every conﬁguration.

Laser 90

◦

Camera 90

◦

Angle Precision Recall F1 Precision Recall F1

◦

0.88 0.91 0.89 0.92 0.94 0.93

◦

0.87 0.89 0.88 0.90 0.93 0.91

◦

0.86 0.90 0.88 0.89 0.92 0.90

detection results of defects. The shape, size, and an-

gle of the defects inﬂuence how the camera will cap-

ture the laser. Figure 6 shows an instant of time t with

the laser on two imperfections that exceed the depth

of the belt in the simulation, with a lagged entry an-

gle (oblique and acute, depending on the position of

the observer/camera) to the belt plane. The laser has

a discontinuous shape due to the hole, and its visu-

alization is only possible while it is reﬂected on the

belt.

Figure 6: Simulation in time t.

Figure 7 presents the result of detecting the points

of high curvature using the statistical analysis to de-

tect outliers. The yellow dots in the image represent

the algorithm’s identiﬁcation.

Figure 7: Outlier detection in simulation database.

As previously stated, detecting defects that cross

the entire belt proﬁle would be the best case for the

proposed method. However, this would clearly be

the worst case in a real environment, as it would be

a serious defect in the belt. However, even in superﬁ-

cial defects, the algorithm could also detect the defect

points, as shown in Figure 8.

The yellow dots describe imperfections on the belt

surface due to their high curvature value. These points

are allocated in a defect matrix after each frame. Af-

ter the laser’s complete passage over any defect, this

matrix will have several points capable of describing

the 2D format of that defect presented in the belt. To

identify these points as a grouping of points (cluster)

describing only one defect at a time, the DBSCAN al-

gorithm was used. Figure 9 shows the defect matrix

after the laser has passed through several defects of

the simulated database.

Figure 10 shows the clustering result using DB-

SCAN algorithm in the defect matrix. It can be seen

that it could group the defects separately, different

bounding boxes, even though the number of defects

Figure 8: Outlier detection on a superﬁcial defect.

Figure 9: Defect matrix.

present in the matrix has not been entered. Also, we

can check the interpretable shape inside the identiﬁed

bounding box. The algorithm input was a vector com-

posed of the indices (row × column) of the defect

matrix points. However, such input data comprises

large and integer values, since they represented pixel

indices in an image. Therefore, they have been nor-

malized for better tuning of the algorithm parameters.

However, as the DBSCAN default parameters have

already managed to identify each defect separately,

these parameters were then used, which are r = 0.1

and n = 10.

Figure 10: DBSCAN cluster in defect matrix.

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

308

Table 2 presents the results of all defect detec-

tion in the simulation. The mean precision was 92%,

which indicates that the proposed method could cap-

ture a high ratio of relevant information for each de-

fect (ratio between TP and FP). The mean recall was

94%, and the mean F1 Score was 91%, indicating

that the method has a high rate of discrimination.

Furthermore, literature shows that good values or a

true detection for F1 are values above 70% (Radau

et al., 2009), and for precision and recall, above 60%

(Genc¸Tav et al., 2012). Thus, we can observe that the

method obtained 100% of detection in the simulated

environment with these values.

Table 2: Results of defect detection using the proposed

technique.

Defect Precision Recall F1

1 0.97 0.91 0.94

2 0.97 0.94 0.96

3 0.86 0.98 0.92

4 0.92 1.00 0.96

5 0.80 0.80 0.80

6 0.93 0.89 0.91

7 0.93 0.97 0.95

8 0.76 0.93 0.84

9 0.92 0.93 0.92

10 0.99 0.89 0.94

11 0.93 0.99 0.96

12 0.90 0.92 0.94

13 0.99 0.98 0.95

14 0.86 0.98 0.92

15 0.98 0.98 0.98

16 0.97 0.95 0.96

mean 0.92 0.94 0.93

6 CONCLUSIONS

The present work proposed a machine vision method

for detecting defects in conveyor belts. A synthetic

database investigated the best conﬁguration between

camera, laser, and belt, as well as the method’s detec-

tion capacity.

The proposed method achieved a good result in

the synthetic environment with a mean precision of

92%, mean recall 94%, and mean F1 score of 93%,

indicating that the method has a good performance

for the identiﬁcation of target defects.

The performance shows that it has, in fact, the

potential for identifying defects in the real environ-

ment of the industry. Moreover, the contribution of

the method to the industry is great since the proposal

is a remote interpretable monitoring system capable

of presenting defects to the maintenance team with-

out the necessity of an operator present at the location

of the conveyor, an environment extremely dangerous

and conducive to accidents, providing a safe and non-

invasive way of monitoring. Also, with the camera

calibration, it would be possible to perform a classiﬁ-

cation only looking at the size of defects, easy deﬁn-

ing a threshold to it.

Future work is intended to assemble the equip-

ment in a real operating situation of a conveyor belt

to construct a database with a more signiﬁcant num-

ber and variability of real defects. This new database

will allow an exhaustive investigation in real situa-

tions and critical parameters for detection. All these

improvements would be valuable, and once the de-

fect is detected, we can develop models capable of

recognizing the type of defect, and even learn how

defect evolves. Machine learning techniques may be

included in the investigation of the evolution of the

defects in such a way as to allow the construction of

predictive models that aid in maintenance.

ACKNOWLEDGEMENTS

This study was ﬁnanced in part by the Coordenac¸

de Aperfeic¸oamento de Pessoal de N

ıvel Superior

- Brasil (CAPES) - Finance Code 001, the Con-

selho Nacional de Desenvolvimento Cient

ıﬁco e Tec-

nol

ogico (CNPQ), the Fundac¸

ao de apoio a pesquisa

do estado de Minas Gerais (FAPEMIG) (FAPEMIG-

APQ-01331-18), the Instituto Tecnol

ogico Vale (ITV)

and Vale S.A. (No.23109.005575/2016-81 and No.

23109.005909/2018-89) and the Federal University of

Ouro Preto (UFOP) (No. 23109.003515/2018-96 and

No.23109.000928/2020-33).

REFERENCES

Elad, M. (2002). On the origin of the bilateral ﬁlter and

ways to improve it. IEEE Transactions on image pro-

cessing, 11(10):1141–1151.

Ester, M., Kriegel, H.-P., Sander, J., Xu, X., et al. (1996).

A density-based algorithm for discovering clusters in

large spatial databases with noise. In Kdd, volume 96,

pages 226–231.

Estrozi, L. F., Rios-Filho, L. G., Bianchi, A. G. C., Ce-

sar Jr, R. M., and da Fontoura Costa, L. (2003). 1d

and 2d fourier-based approaches to numeric curvature

estimation and their comparative performance assess-

ment. Digital signal processing, 13(1):172–197.

Fedorko, G., Moln

ar, V., Dovica, M., T

oth, T., Fabianov

J., Strohmandl, J., Neradilov

a, H., Heged

s, M., and

Belu

sko, M. (2018). Analysis of defects in carcass of

rubber–textile conveyor belts using metrotomography.

Journal of Industrial Textiles, 47(7):1812–1829.

Fedorko, G., Molnar, V., Marasova, D., Grincova, A.,

Dovica, M., Zivcak, J., Toth, T., and Husakova, N.

(2014). Failure analysis of belt conveyor damage

caused by the falling material. part i: Experimental

measurements and regression models. Engineering

failure analysis, 36:30–38.

Early Defect Detection in Conveyor Belts using Machine Vision

309

Fromme, C., Stager, D., Pilarski, T., Bancroft, B., and

Hegadorn, T. (2006). Conveyor belt inspection sys-

tem and method. US Patent 6,988,610.

fu Hou, Y. and rui Meng, Q. (2008). Dynamic characteris-

tics of conveyor belts. Journal of China University of

Mining and Technology, 18(4):629 – 633.

Genc¸Tav, A., Aksoy, S., and

ONder, S. (2012). Unsuper-

vised segmentation and classiﬁcation of cervical cell

images. Pattern recognition, 45(12):4151–4168.

Guan, Y., Zhang, J., Shang, Y., Wu, M., and Liu, X.

(2008). Embedded sensor of forecast conveyer belt

breaks. In Fuzzy Systems and Knowledge Discovery,

2008. FSKD’08. Fifth International Conference on,

volume 5, pages 617–621. IEEE.

Hodge, V. and Austin, J. (2004). A survey of outlier de-

tection methodologies. Artiﬁcial intelligence review,

22(2):85–126.

Jeong, J., Park, E., Han, W. S., Kim, K., Choung, S., and

Chung, I. M. (2017). Identifying outliers of non-

gaussian groundwater state data based on ensemble

estimation for long-term trends. Journal of Hydrol-

ogy, 548(Supplement C):135 – 144.

Junior, R. M. C. and Costa, L. D. F. (1996). Towards effec-

tive planar shape representation with multiscale digi-

tal curvature analysis based on signal processing tech-

niques. Pattern Recognition, 29(9):1559–1569.

Kurihara, T., Okuda, K., and Yoshimoto, M. (2006). Longi-

tudinal rip detection method and device for conveyor

belt. Japan Patent 6,282,319.

Leys, C., Ley, C., Klein, O., Bernard, P., and Licata, L.

(2013). Detecting outliers: Do not use standard devi-

ation around the mean, use absolute deviation around

the median. Journal of Experimental Social Psychol-

ogy, 49(4):764–766.

Li, J. and Miao, C. (2016). The conveyor belt longitudi-

nal tear on-line detection based on improved ssr al-

gorithm. Optik-International Journal for Light and

Electron Optics, 127(19):8002–8010.

Li, M., Du, B., Zhu, M., and Zhao, K. (2011). Intelligent de-

tection system for mine belt tearing based on machine

vision. In Control and Decision Conference (CCDC),

2011 Chinese, pages 1250–1253. IEEE.

Lin, L. and Sherman, P. D. (2007). Cleaning data the chau-

venet way. The Proceedings of the SouthEast SAS

Users Group, SESUG Proceedings, Paper SA11.

Nascimento, R., Carvalho, R., Delabrida, S., Bianchi, A. G.,

Oliveira, R. A. R., and Garcia, L. G. U. (2017). An in-

tegrated inspection system for belt conveyor rollers. In

19th International Conference on Enterprise Informa-

tion Systems (ICEIS 2017), volume 2, pages 190–200.

Nicolay, T., Treib, A., and Blum, A. (2004). Rf identiﬁ-

cation in the use of belt rip detection [mining product

belt haulage]. In Sensors, 2004. Proceedings of IEEE,

pages 333–336. IEEE.

Otsu, N. (1979). A threshold selection method from gray-

level histograms. IEEE transactions on systems, man,

and cybernetics, 9(1):62–66.

Pang, Y. and Lodewijks, G. (2005). Large-scale conveyor

belt system maintenance decision-making by using

fuzzy causal modeling. In Intelligent Transportation

Systems, 2005. Proceedings. 2005 IEEE, pages 563–

567. IEEE.

Pang, Y. and Lodewijks, G. (2006). A novel embedded con-

ductive detection system for intelligent conveyor belt

monitoring. In Service Operations and Logistics, and

Informatics, 2006. SOLI’06. IEEE International Con-

ference on, pages 803–808. IEEE.

Peng, X. (2013). A novel image-based method for conveyor

belt rip detection. In Signal Processing, Communica-

tion and Computing (ICSPCC), 2013 IEEE Interna-

tional Conference on, pages 1–4. IEEE.

Perruchet, P. and Peereman, R. (2004). The exploitation

of distributional information in syllable processing.

Journal of Neurolinguistics, 17(2-3):97–119.

Qiao, T., Li, X., Pang, Y., L

u, Y., Wang, F., and Jin, B.

(2017). Research on conditional characteristics vision

real-time detection system for conveyor belt longitu-

dinal tear. IET Science, Measurement & Technology,

11(7):955–960.

Qiao, T., Liu, W., Pang, Y., and Yan, G. (2016). Research

on visible light and infrared vision real-time detection

system for conveyor belt longitudinal tear. IET Sci-

ence, Measurement & Technology, 10(6):577–584.

Radau, P., Lu, Y., Connelly, K., Paul, G., Dick, A., and

Wright, G. (2009). Evaluation framework for algo-

rithms segmenting short axis cardiac mri.

Sasaki, Y. et al. (2007). The truth of the f-measure. Teach

Tutor mater, 1(5):1–5.

Xianguo, L., Lifang, S., Zixu, M., Can, Z., and Hangqi, J.

(2018). Laser-based on-line machine vision detection

for longitudinal rip of conveyor belt. Optik, 168:360–

369.

Yang, Y., Miao, C., Li, X., and Mei, X. (2014). On-

line conveyor belts inspection based on machine vi-

sion. Optik-International Journal for Light and Elec-

tron Optics, 125(19):5803–5807.

Yang, Y., Zhao, Y., Miao, C., and Wang, L. (2016). On-line

longitudinal rip detection of conveyor belts based on

machine vision. In Signal and Image Processing (IC-

SIP), IEEE International Conference on, pages 315–

318. IEEE.

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

310