Journal of Food Engineering 83 (2007) 384393

www.elsevier.com/locate/jfoodeng

Citrus sorting by identication of the most common defects

using multispectral computer vision
J. Blasco a,*, N. Aleixos b, J. Gomez a, E. Molto a
a

Centro de AgroIngeniera, Instituto Valenciano de Investigaciones Agrarias (IVIA), Ctra. Moncada-Naquera km 5, 46113 Moncada (Valencia), Spain
b
Departamento de Ingeniera Mecanica y Construccion, Universitat Jaume I de Castello (UJI), Campus Riu Sec, 12071 Castellon, Spain
Received 5 October 2006; received in revised form 10 March 2007; accepted 12 March 2007
Available online 24 March 2007

Abstract
The presence of skin defects is one of the most inuential factors in the price of fruit. The detection of defects during packing operations ensures that only fruits with a good quality reach the market. Moreover, the identication of the type of each defect will increase
both the quality of the fruit and also the producers prot. At the present time, fruit with slight defects is marketed together with sound
fruit, thus depreciating the quality of the batch, or it is removed together with seriously damaged fruit, thereby causing economic losses.
Most current computer vision systems used in the automatic quality inspection of food are limited to the visible region of the electromagnetic spectrum as they tend to imitate the human eye. However, non-visible information, such as that provided by near-infrared or
ultraviolet regions of the spectrum, can improve the inspection by detecting specic defects or allowing the detection of non-visible damages. This work summarises our research in the application of near-infrared, ultraviolet and uorescence computer vision systems in the
identication of the most common defects of citrus fruits, and proposes a fruit sorting algorithm that combines this dierent spectral
information (including visible) to classify fruit according to the type of defect. Results showed that the contribution of non-visible information can improve the detection and identication of some defects. Compared with the results from colour images, the detection accuracy of anthracnose increased from 86% by using NIR images; and the accuracy of green mould was increased from 65% to 94% by using
images of uorescence.
2007 Elsevier Ltd. All rights reserved.
Keywords: Fruit sorting; Machine vision; Multispectral imaging; Image analysis; Fruit quality

1. Introduction
The presence of skin defects is one of the most inuential factors in the quality and price of fresh fruit, since most
consumers associate quality with a good appearance and
the total absence of external defects. For this reason, packinghouses demand more complex systems that are capable
of detecting skin defects without misrecognising them as
the fruit stem. The importance of detecting defects and
their connection with the quality can be seen by the number
of research publications dealing with machine vision systems or image processing techniques for detecting the
*

Corresponding author. Tel.: +34 963424000; fax: +34 963424001.

E-mail address: [email protected] (J. Blasco).

0260-8774/$ - see front matter 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jfoodeng.2007.03.027

defects of dierent types of fruits such as peaches (Miller

& Delwiche, 1991), olives (Diaz et al., 2004), oranges (Cerruto, Failla, & Schillaci, 1996), potatoes (Muir, Porteus, &
Wastie, 1982), bell peppers (Shearer & Payne, 1990), stonefruit (Singh & Delwiche, 1994), pistachio (Pearson, 1995),
dry dates (Wulfsohn, Sarig, & Algazi, 1993), sweet cherry
(Guyer, Uthaisombut, & Stockman, 1996), and mushroom
(Heinemann et al., 1994). In the same way, due to the economic importance of the crop, apples have been the subject
of a number of studies (Leemans & Destain, 2004; Li,
Wang, & Gu, 2002; Wen & Tao, 1999).
Nowadays, the advances in the latest processors and the
lower costs of modern computers allow a great amount of
data to be processed in a short time. Consequently, bigger
images can be analysed using more complex algorithms

J. Blasco et al. / Journal of Food Engineering 83 (2007) 384393

that allow the detection of external defects while working

on-line (Diaz, Faus, Blasco, Blasco, & Molto, 2000). However, beyond the detection of defects, it is also very important for producers to know which type of defects are
aecting each fruit so they can separate rotten or seriously
damaged fruits which must be rejected, from other fruits
containing slight damages that only aects their appearance and can be marketed as second quality (Blasco &
Molto, 2002). Another benet to be gained from knowing
what sort of defects or diseases are aecting the fruits will
allow the growers to apply treatments more accurate, thus
saving chemical products with the ensuring economical and
environmental benets.
Since the colour is the natural sense we use to make our
rst evaluation of the quality of fruits, most of the inspection systems use only this feature to segment the images
(Leemans, Magein, & Destain, 1998). Several works performed in this way succeed in detecting the external defects
(Blasco, Aleixos, & Molto, 2003; Leemans, Magein, &
Destain, 1999) although most of them do not distinguish
between dierent types of defects, neither the confusion
between the stem nor the calyx and the true defects. The
main problem of the techniques based on the pixel inspection in colour images, is related to the natural colour variation of biological produce that complicates the
detection of objects with changing colours such as defects,
since the same type of defect can present dierent colours
in the same fruits. In the case of the stem or calyx, specic
research is therefore needed to correctly identify them (Penman, 2001; Ruiz, Molto, Juste, Pla, & Valiente, 1996).
Unlike the human eye, machine vision is not limited to
the visible (VIS) region of the electromagnetic spectrum
and several studies have been carried out to inspect fruits
using ultraviolet (UV) or near-infrared (NIR) radiation
(Alchanatis, Peleg, & Ziv, 1993; Rehkugler & Throop,
1986). Results show that the combination of NIR and
VIS can improve the identication of defects in fruits (Aleixos, Blasco, Navarron, & Molto, 2002; Steinmetz, Roger,
Molto, & Blasco, 1999). The sensitivity of cameras based
on CCD technologies is limited to the range between
approximately 400 and 1000 nm, and so spectrometry has
been widely used to investigate the importance of NIR
(Abu-Khalaf, Bennedsen, & Bjrn, 2004; Blasco, Ortiz,
Sabater, & Molto, 2000; Peirs, Lammertyn, Nicola, &
De Baerdemaeker, 2000; Rodriguez-Otero, Hermida, &
Ceneno, 1997; Schulz, Drews, Quilitzsch, & Kruger,
1998; Zude, 2003).
Another widely extended technique that can be applied
to detect invisible damage is based on uorescence since
essential oils released by the damaged glands are uorescent
(Ariana, Guyer, & Shrestha, 2004; Skjervold et al., 2003;
Vargas, Kim, Tao, Lefcourt, & Chen, 2004). For this reason, many citrus packinghouses have a special closed room,
illuminated with ultraviolet light, for the detection of rottenness which is mostly caused by fungi as penicillium digitatum (green mould). The conveyor belts transporting the
fruit cross these rooms, where the operators must wear pro-

385

tective glasses and gloves because long exposure to ultraviolet radiation can be dangerous for human skin. The
technique used for detecting the green mould is uorescence. On being lit by ultraviolet light, the rottenness produces uorescence that can be detected by the operators.
Multispectral inspection systems can be created by the
combination of systems that are sensitive to dierent
regions of the spectrum. The application of these multispectral systems in food safety and inspection can be
applied to the detection of some features normally aecting
the quality (Chun-Chieh, Kuanglin, & Yud-Red, 2005;
Kleynen, Leemans, & Destain, 2005; Mehl, Chao, Kim,
& Chen, 2002; Mehl, Chen, Kim, & Chan, 2004; Yang,
Chao, & Chen, 2005). Following on from this, our work
proposes a multispectral inspection system that combines
four image acquisition systems that are sensitive to the visible and non-visible regions of the spectrum. Then, the
multispectral information was used to develop a fruit sorting algorithm based on the identication of the most common external defects of oranges and mandarins.
2. Objective
The aim of the work is to develop a system to sort citrus
fruits according to the type of external defects they present.
Four steps had to be fullled in order to achieve this objective: (1) to select a colour space with highest discriminatory
power to classify the dierent blemishes in the skin of citrus
fruits, (2) to detect and classify defects in citrus fruits using
near-infrared, visible, ultraviolet and uorescence, (3) to
develop a fruit sorting algorithm that combines the multispectral (including visible) information according to the
type of external defects, and is therefore capable of distinguishing between slight and serious damage and (4) to demonstrate that the multispectral system improves the
performance of a system based only in visible imaging.
3. Materials and methods
3.1. Image acquisition
The colour image acquisition system used in the tests
consisted of a three CCD (coupled charge device) camera
(Sony XC-003P) capable of acquiring images with a size
of 768  576 pixels. The lighting system was composed
by eight uorescent tubes (daylight type, 25 w) and polarised lters in order to avoid bright spots in the scene by
cross-polarisation. Images containing one fruit, with a resolution of 0.17 mm/pixel, were acquired o-line. The side
of the fruit that contained the defects being presented manually to the camera. If necessary, several images were
acquired for each fruit.
The uorescence images were acquired using the same
colour camera but another illumination system. This lighting system was composed of eight uorescent tubes of
black light (18 w each) that emit radiation with a wavelength between 350 and 400 nm, with the peak at 370 nm.

386

J. Blasco et al. / Journal of Food Engineering 83 (2007) 384393

Fig. 1. Inspection chambers used in the acquisition of the fruit images.

Following the work described in Ruiz, 1999, the images

were acquired by placing a 560 nm interferometric lter
on the camera lens.
The acquisition of the NIR images was performed using
a Hamamatsu BeamFinder III C5332-01 camera, based on
a vidicon tube that is sensitive from 400 nm to 1800 nm.
The output of the camera is composite video and the
images had a size of 768  576 pixels and a resolution of
0.18 mm/pixel. The lighting system was composed of two
incandescent lamps (Philips R125-IR, 250 w). To avoid
the interference of visible information, a 700 nm cut-band
lter was coupled to the camera lens.
The camera employed for the acquisition of the UV
images was a Proxitronic Proxicam HL1, which is based
on CCD technology in combination with an image intensier. This camera can acquire images with a size of
756  581 pixels in the spectral range between 200 and
800 nm. The images were acquired with a resolution of
0.18 mm/pixel. The lighting system was the one used to
acquire the uorescence images. To avoid the inuence of
visible radiation a 400 nm cut-band lter was placed on
the camera optics. The cameras were mounted on a conveyor belt similar to those used in the fruit inspection lines.
In all the cases, the damaged part of the fruits was presented to the camera manually. In one imaging season. The

same fruits were imaged with the four systems to avoid differences in the detection of the damages caused by the time
factor. The images were acquired used the inspection
chambers shown in Fig. 1. Fig. 2 shows a scheme of the
cameras and the lighting systems distribution that were
placed into the inspection chambers.
The experiments were carried out using images of 428
defects in oranges (navelina and valencia varieties) and
369 defects in mandarins (marisol, clemenules and fortune
varieties) provided and labelled by the biological analysis
department of a manufacturer company of automatic
inspection machines. All the fruits were manually analysed
using microscopes, and the defects found were identied,
labelled and numbered. Four images of each fruit were
acquired (visible, near-infrared, ultraviolet and uorescence), thus providing a collection of 3188 images (797
images acquired using each technique). The images represent 162 stems and 635 labelled defects, containing each
image contain only one type of defect.
The defects were divided into serious and slights. Serious defects were those that cause the loss of the fruit or
that can spread infestations to other fruits in the packing
house: anthracnose, stem-end injury, green mould and
medy egg deposition. On the other hand, defects were
considered as slight when they aect only the appearance
of the fruit and include oleocellosis (rind oil spot), presence of scales, scarring, thrips, chilling injury, sooty
mould and phytotoxicity. Table 1 details the type and distribution of the dierent defects in the tested fruits. Fig. 3
shows dierent scenes of a fruit aected by green mould
acquired using the dierent cameras. It appears in dierent positions because the fruit was imaged using cameras
placed in dierent inspection chambers, so moved from
one image to other but presenting the same defect.
Another 162 fruits contained no defects but, in these
cases, images of the stem were acquired.
3.2. Selection of the colour space for the analysis in the
visible
The classication of the regions of interest in the colour
images, such as stems or specic defects, was based on the

Fig. 2. Schemes of the vision systems showing the cameras and lamps positions.

J. Blasco et al. / Journal of Food Engineering 83 (2007) 384393

387

Table 1
Distribution of the defects and stem in the images of the fruit used for the experiments
Oranges
Navelina

Mandarins
Valencia

Marisol

Total defects
Clemenules

Fortune

Slights defects

Thrips
Phytotoxicity
Scarring
Scale
Chilling injury
Sooty mould
Oleocellosis

76
10
50
11
6
14
30

18
7
24
8
5
8
14

20
0
15
13
0
7
15

16
0
9
5
0
5
13

25
6
28
7
0
0
13

155
23
126
44
11
34
85

Serious defects

Anthracnose
Stem-end injury
Medy
Green mould

6
0
27
16

5
9
0
12

6
11
9
11

0
11
5
6

5
8
0
10

22
39
41
55

Stem

40

32

33

24

33

162

Total

286

142

140

94

135

797

Fig. 3. The same fruit aected by green mould imaged in (from left to right) visible, near-infrared, uorescence and ultraviolet.

colour information, so the rst step was to select the most

discriminating colour space. The colour spaces studied
were RGB, XYZ, HSI, La*b* and Lu*v*. The images were
acquired in RGB coordinates because it is the native space
for colour representation in computers. To perform the
comparison, 15 training images containing a representative
selection of all types of defects and the stems were selected
by an expert. A semiautomatic application was programmed to manually select representative pixels belonging
to each type of defect and the stems, storing their colour
coordinates together with a label describing the type of
defect. The colour coordinates were also transformed from
RGB to the rest of the colour spaces before using the colour space conversion equations described in Poynton
(1995, 1997).
Linear discriminant analysis (LDA) was used to obtain
classication functions in each colour space, using the colour coordinates as independent variables. A total of 12
classication functions (corresponding to eleven defects
and stem) were calculated for each colour space using a statistical computer application. The classication functions
were similar to the described in the point 3.3. A set of
images, dierent from the used for the experiments, with
known types of defects and presence of stems was classied
with these functions and the average success rate in the

classication obtained in the dierent spaces were then

compared. The colour space with the highest success rate
was selected for analysing the visible images.
3.3. Classication of defects in using visible colour images
The performance of the system based only in visible
images alone was studied for comparison with the multispectral procedure proposed. In the case of visible images,
the result of the segmentation algorithm consists in regions
of interest in each image that can be blemishes or the stem,
but without providing any additional information. To classify these regions, segmented images were used to select,
label and store representative pixels of stems and of each
type of blemish (obtained from training samples). In dierent steps, pixels belonging to the same segmented region of
interest were selected using the mouse of the computer and
assigned (labelled) manually to a particular class of defect.
This process was repeated for all the defects found in the
training set of images. The colour coordinates of these pixels in the original image were used as independent variables
to perform a discriminant analysis, being the label with the
type of the defect used as grouping variable. Classication
functions were calculated for each class in the model. The
classication functions can be used to determine to which

388

J. Blasco et al. / Journal of Food Engineering 83 (2007) 384393

group each case most likely belongs. There are as many

classication functions as there are groups (12 in our case).
Each function allows us to compute classication scores for
each case for each group (StatSoft, 2003), by applying the
Eq. (1). The classication functions were computed for
each group and used to classify the cases (defects). We
would classify a case into the group for which it has the
highest classication score.
Z j wHj  H wSj  S wIj  I cj

where Zj is the classication score for the class j. The variables H, S, I, are the coordinates of the averaged colour of
the sample, wHj ; wSj ; wIj are weights that represent the
importance of each variable for each class and Cj is a constant that is dierent for each class. The variable j
(j = 1, . . ., 12) is the class and represent the eleven types of
the studied defects plus the stem.

Fig. 4. NIR images of fruits aected by anthracnose (left) and sooty

mould (right).

3.4. Detection of defects in NIR, UV and FL

One of the aims of this research was to evaluate the
importance of non-visible spectral information in identifying types of defects present in the fruits. The strategy was
based on distinguishing the defects in colour images and
then improving their identication using the contribution
of the spectral information provided by NIR, UV and uorescence images, without calculating morphological features of the defects. The segmentation of the visible and
uorescence images was carried out using the technique
described in Blasco, Aleixos, and Molto (2007) which is
based in an unsupervised region growing process, selecting
the initial set of seeds in the more homogenous regions of
the image and adding pixels to the seeds until complete
the entire image. To process the NIR and the UV images,
the same procedure was employed but adapted for working
with grey-level values instead of colour coordinates.
Illuminating with a radiation below the 400 nm and
avoiding the inuence of external light, the only radiation
than can reach the CCD of the camera using that lter,
is the caused by the uorescence phenomenon induced in
the essential oil of the peel of the fruits when it is released
by some defects, as rottenness (Latz & Ernes, 1978). As a
consequence, only these defects could be detected using this
kind of system. These technique is commonly using in the
packing lines to detect rotten fruit and remove it from
the line manually.
The NIR imaging is commonly employed, in combination of visible imaging, in the commercial electronic fruit
sorters based on machine vision, to discriminate between
fruits and background, since the skin of the fruits reects
almost all the NIR infrared while the background absorbs
it.
The experiments showed that only anthracnose and
sooty mould were detected in NIR (Fig. 4). In the same
way, only the stem-end injury was detected in UV
(Fig. 5) and, only thrips, scarring and green mould were
detected in the FL images. However, an important nding

Fig. 5. UV image of a fruit aected by stem-end injury.

Fig. 6. FL image of a fruit aected by green mould (left) and thrips

(right).

was that no false detections were observed in these images.

Fig. 6 shows the image of fruits aected by green mould
and thrips converted to grey-level during a process of the
detection of the defects. This information led us to design
of the algorithm described in Fig. 7 for classifying the fruit
according to the severity of the damage.
3.5. Fruit sorting using a multispectral inspection system
Both, visible and non-visible information are combined
in this algorithm to obtain a sorting system of citrus fruits.
The citrus classication algorithm proposed (Fig. 7) makes
prior the detection of those defects that are more dangerous for the commercialisation of the fruits.

J. Blasco et al. / Journal of Food Engineering 83 (2007) 384393

Fig. 7. Flowchart of the algorithm for classifying the fruit according to the severity of the damage.

389

390

J. Blasco et al. / Journal of Food Engineering 83 (2007) 384393

1. The rst images to be analysed were the UV because the

only one type defect that could be sensed in this type of
image was a serious one (stem-end injury). If any defect
was found in the UV image, the fruit was directly classied as defected by stem-end injury and the algorithm
nishes for this fruit.
2. If not defects are found in the UV image, the algorithm
follows with the analysis of the colour images searching
for other serious defects. If any defect is detected after
the segmentation, it is classied using the classication
function. The classication score Zj is obtained for all
type of studied defects. If the defect with the highest
classication score is any serious defect, the fruit is classied as containing that defect and the algorithm ends
for this fruit. In this case the potential serious defects
are anthracnose, medy or green mould.
3. If not defects have been detected in this point, the next
image to be analysed is the NIR image. If any defect
was found in the NIR images, it was supposed to be
anthracnose (could be a fruit aected by anthracnose
that was not detected in the previous step) or sooty
mould since only this kind of defects could be detected
using NIR in our experiments. Since the decision
between these defects cannot be taken using only the
NIR information because it does not discriminate
between them, the defect is classied attending to the
classication score obtained by each defect during the
inspection of the visible image in the previous step. If
ant defect is found in the NIR image, the algorithm ends
for this fruit, otherwise it continues with the analysis of
the uorescence image.
4. The last step consisted in processing the uorescence
images in the same way as the NIR images. Only thrips,
scarring or green mould were found in these type images
but, as in the case of NIR, the system could not discriminate between them. So if any defect is found using this
technology, again the classication score obtained by
these defects when analysing the visible image was used
to discriminate between these tree defects.
5. If no defects have been found using UV, NIR or FL,
then, the visible image is asked again to check if any
other damage was found on it. If so, since all the serious
defects have been already detected in the previous processes, they only can be light damages. The fruit is classied as containing the damage with the highest
classication score calculated in step 2.
6. If no defects have been encountered at this point, the
fruit is classied as sound fruit.

Table 2
Classication of blemishes and stem in dierent colour spaces
Colour space

Total

Average of correct classication in each colour space (%

success)
HSI

La*b*

RGB

Lu*v*

XYZ

87.2

83.3

83.7

82.1

71.3

fying 87.2% of defects. Although the results have been similar for all the space colours (but XYZ), HSI is a non-linear
deformation of the RGB cube, the higher success rate
owes, probably, because it represents better the small colour variability in the samples (orange of the skin, grey or
brown of some defects and green of stems). This result
agrees with the obtained by Du and Sun (2005) who compared dierent colour space transformations using dierent
algorithms. Although they employed the HSV instead HSI,
both transformations are similar. Consequently, this colour
space was used to generate the statistical model for identifying the defects in the visible images.
Regarding the defect classication the results were
obtained analysing the images using the algorithm proposed and comparing the classication performed by the
algorithm with the actual damage of the fruit. The results
in tables are expressed as percentage of success rate following Eq. (2).
Success rate number of fruits correctly classified  100
 number of total damaged fruits1

Oils are mixtures which often contain compounds that

uoresce, this is the case of the essential oils contained in
the skin of the citrus fruits. Some damages can release these
essential oils, allowing the detection of the defects by
means of the uorescence principle. Table 3 shows the result of the uorescence image analysis. Only three dierent
types of defects were detected, in percentage of fruit detection ranging from 50%, in the case of scarring, to 93% in
the case of green mould. However, the system was unable
to dierentiate between them.
Table 4 shows the results of the analysis of the UV
images. The stem-end injury was the only defect detected.
This defect is a consequence of the absence of skin around
the stem area and thus represents a way for fungus and
other pathogens. The skin of the citrus absorbs all the

Table 3
Performance of uorescence image analysis

4. Results and discussion

Type of
defect

Total
fruits

Table 2 shows the results of the study of the colour

spaces for our experiments. Although similar results were
obtained using the dierent colour space transformations,
the greatest discriminatory power corresponded to the
HSI colour space, which was capable of successfully classi-

Green
mould
Scarring
Thrips

55

Total

Detected
fruits

False
detections

Success rate
(%)

51

92.7

126
155

63
99

0
0

50.0
63.9

336

213

63.4

J. Blasco et al. / Journal of Food Engineering 83 (2007) 384393

Table 4
Performance of ultraviolet image analysis
Type of
defect

Total
fruits

Detected
fruits

False
detections

Success rate
(%)

Stem-end
injury

39

31

79.5

Table 5
Performance of near-infrared image analysis
Type of
defect

Total
fruits

Detected
fruits

False
detections

Success rate
(%)

Sooty
mould
Anthracnose

34

31

91.2

22

21

95.4

Total

56

52

92.9

UV radiation but, in these areas where the skin has been

removed, as in the case of the stem-end injury, the UV light
is reected allowing the detection of the defects. The success rate of stem-end injury detection was 80%, and again
no false detections were observed.
Results from the NIR images are shown in Table 5.
Only two kinds of defects were detected: sooty mould
(slight defect) and anthracnose (serious defect). As in the
case of uorescence images, both defects could not be discriminated but 93% of them were detected. Since no false
detections were obtained, if any defect was found in this
type of images, it should be one of the two types of previously determined defects. This information was employed
to enhance identication using colour images. It is very
important that the serious damage found in this type of
images, such as anthracnose, was detected in 100% of cases,
thus ensuring the correct classication of the fruit as damaged. The skin of biological produces, as fruits, reects the
NIR radiation. However, some defects can alter this property as the case of the anthracnose, that causes the disinte-

391

gration of the external tissues of the peel, or the sooty

mould, that stains the peel absorbing the radiation.
Although the success rates using these spectral images
are not high, it is very important to remark the fact that
no false detections were obtained using these techniques
in the conditions under which the tests were conducted
and while studying these particular defects. This means
that no sound fruits were detected as defected and that if
a defect is found in this type of images, this defect should
be one of the three types that appear in tables. These
results, however, do not rule out possible confusion with
other type of defects not studied in this work, but which
are normally more dicult to nd in real conditions.
Table 6 shows the results of classifying fruit using only
visible information. The rst column represents the type
of defect. The rest of the columns correspond to how each
fruit was classied. Most defects were correctly detected
although, in general, the percentage of success was low.
It is interesting to note the high performance achieved in
recognising fruits aected by particular defects as chilling
injury (100%), sooty mould (82%) or anthracnose (86%)
while the identication of the others, such as, green mould
(65%) or medy (62%), was low.
Table 7 shows the results using the multispectral decision algorithm described. Compared with the results
obtained from only the colour images, it is obvious that
the inuence of non-colour images increases the rate of success in fruit classication, especially in the case of seriously
damaged fruits, which are the most important fruits that
must be removed from the market line. Using the NIR
information, the identication of fruits with anthracnose
increased from 86% to 95%. With uorescence images,
the properly classied rotten fruit aected by green mould
rose to 95% from 65% using only colour information; slight
damage, such as thrips or scarring, increased to 87% and
81% from 78% and 71%, respectively. However, the most
successful identication was achieved in fruits with stem-

Table 6
Fruit classication according to the damage identied using only colour images (slight AG, serious HK)
Detected damages

Actual damage

Stem

A Thrips
B Phytotoxicity
C Scarring
D Scale
E Chilling injury
F Sooty mould
G Oleocellosis

121
0
21
0
11
0
0

4
15
0
0
3
0
0

25
1
90
0
10
0
0

13
1
4
19
3
0
0

0
0
0
0
11
0
0

0
0
0
0
0
28
0

2
3
2
0
8
0
48

2
0
0
0
0
1
0

0
0
0
0
0
0
0

0
2
0
0
0
0
7

1
0
1
0
3
0
11

0
0
1
0
0
0
0

H Anthracnose
I Stem-end injury
J Medy
K Green mould

0
0
2
0

0
0
1
0

0
0
0
0

3
0
1
0

0
0
0
0

6
0
0
0

0
0
8
14

19
0
0
0

0
36
0
3

6
0
25
0

0
0
3
36

0
0
0
0

Stem (no damage)

12

21

12

108

Total

155

23

126

44

11

34

85

22

39

40

55

162

Success rate (%)

78.1

65.2

71.4

43.2

100

82.4

56.5

86.4

92.3

62.5

65.5

66.7

392

J. Blasco et al. / Journal of Food Engineering 83 (2007) 384393

Table 7
Fruit classication using the multispectral algorithm proposed (slight AG, serious HK)
Detected damages

Actual damage
A

A Thrips
B Phytotoxicity
C Scarring
D Scale
E Chilling injury
F Sooty mould
G Oleocellosis

135
0
15
0
5
0
0

4
15
0
0
3
0
0

19
1
102
0
4
0
0

13
1
4
19
3
0
0

0
0
0
0
11
0
0

0
0
0
0
0
28
0

2
3
2
0
8
0
48

0
0
0
0
0
1
0

0
0
0
0
0
0
0

0
2
0
0
0
0
7

0
0
0
0
0
0
2

0
0
1
0
0
0
0

H Anthracnose
I Stem-end injury
J Medy
K Green mould

0
0
0
0

0
0
1
0

0
0
0
0

3
0
1
0

0
0
0
0

6
0
0
0

0
0
8
14

21
0
0
0

0
39
0
0

6
0
25
0

0
0
1
52

0
0
0
0

Stem (no damage)

12

21

12

108

155

23

126

44

11

34

85

22

39

40

55

162

82.4

56.5

95.5

62.5

94.6

Total
Success rate (%)

87.1

65.2

Stem

81.0

43.2

100

end injury, which rose from 92% to 100% of the fruits correctly identied and classied. The rest of the defects did
not undergo any kind of modication in the percentage
of identication because they were not detected in the rest
of the images.
These results could possibly be improved by introducing
morphological information, such as the size or shape
descriptors, into the algorithms for the identication of
the defects. This is the case of scale, which is a defect often
confused with thrips, however, with much smaller size compared to the latter.
5. Conclusions
Five dierent colour spaces were compared in order to
obtain the most appropriate for classifying the dierent
defects studied and the stems. The results of discriminant
analysis were similar with the dierent colour spaces but
the highest success rate (87%) was achieved with HSI.
This research contributes to evaluate of the NIR, UV
and FL as potential regions to the correct identication
of several external defects aecting citrus, especially some
of the most dangerous defects, as the green mould caused
by penicillium digitatum. With the UV system only the
stem-end injury was detected with no false detections,
which means that all the detections corresponded to true
stem-end damages. Using NIR images, only anthracnose
and sooty mould were detected, while uorescence images
enabled us to detect damages caused by green mould, scarring or thrips. As in the UV, no false detections were
detected with these systems, but it was not possible to distinguish between the defects.
A fruit sorting algorithm that combines visible, NIR,
UV and FL information was proposed. The UV information was important in improving the detection of stemend injury up to 100% of the cases, while the contribution
of NIR and FL played an important role by increasing the

100

66.7

success rate of anthracnose and green mould detection up

to around 95%, all of which are considered to be serious
damage.
Acknowledgements
This work was partially funded by the Spanish Ministry
of Science and Technology (MCYT) by means of the Plan
Nacional de Investigacion Cientca, Desarrollo e Innovacion Tecnologica (I+D+I) and European FEDER
funds, through Project DPI-2003-09173-C02-02 Tecnicas
avanzadas de vision por computador para el reconocimiento e identicacion automatica de los defectos externos
de los ctricos.
References
Abu-Khalaf, N., Bennedsen, B., & Bjrn, G. (2004). Distinguishing
carrots characteristics by near-infrared (NIR) reectance and multivariate data analysis. In: Agricultural engineering international: The
CIGR Journal of Scientic Research and Development. Manuscript FP
03 012.
Alchanatis, V., Peleg, K., & Ziv, M. (1993). Classication of tissue culture
segments by colour machine vision. Journal of Agricultural Engineering
Research, 55, 299311.
Aleixos, N., Blasco, J., Navarron, F., & Molto, E. (2002). Multispectral
inspection of citrus in real-time using machine vision and digital signal
processors. Computers and Electronics in Agriculture, 33(2), 121137.
Ariana, D. P., Guyer, D. E., & Shrestha, B. P., (2004). Integrating
reectance and uorescence imaging for apple disorder classication.
In: Proceedings of the SPIE, monitoring food safety, agriculture, and
plant health. Vol. 5271 (pp. 301312), Boston, USA.
Blasco, J., Aleixos, N., & Molto, E. (2003). Machine vision system for
automatic quality grading of fruit. Biosystems Engineering, 85(4),
415423.
Blasco, J., Aleixos, N., & Molto, E. (2007). Computer vision detection of
peel defects in citrus by means of a region oriented segmentation
algorithm. Journal of Food Engineering, 81(3), 535543.
Blasco, J., & Molto, E. (2002). Identication of defects in citrus skin using
multispectral imaging. In: International conference on agricultural

J. Blasco et al. / Journal of Food Engineering 83 (2007) 384393

engineering, AgEng 02, EurAgEng Paper No. 02-AE-031, Budapest,
Hungary.
Blasco, J., Ortiz, C., Sabater, M. D., & Molto, E. (2000). Early detection
of fungi damage in citrus using NIR spectroscopy. In: Proceedings of
the SPIE, biological quality and precision agriculture II. Vol. 4203 (pp.
4754), Boston, USA.
Cerruto, E., Failla, S., & Schillaci, G. (1996). Identication of blemishes
on oranges. In: International conference on agricultural engineering,
AgEng 96, EurAgEng Paper No. 96F-017, Madrid, Spain.
Chun-Chieh, Y., Kuanglin, C., & Yud-Red, C. (2005). Development of
multispectral image processing algorithms for identication of wholesome, septicemic and inammatory process chickens. Journal of Food
Engineering, 69(2), 225234.
Diaz, R., Faus, G., Blasco, M., Blasco, J., & Molto, E. (2000). The
application of a fast algorithm for the classication of olives by
machine vision. Food Research International, 33, 305309.
Diaz, R., Gil, L., Serrano, C., Blasco, M., Molto, E., & Blasco, J. (2004).
Comparison of three algorithms in the classication of table olives by
means of computer vision. Journal of Food Engineering, 61(1),
101107.
Du, C. J., & Sun, D. W. (2005). Comparison of three methods for
classication of pizza topping using dierent colour space transformations. Journal of Food Engineering, 68(3), 277287.
Guyer, D., Uthaisombut, P., & Stockman, G. (1996). Tissue reectance
and machine vision for automated sweet cherry sorting. In: Proceedings of the SPIE, optics in agriculture, forestry, and biological processing
II. Vol. 2907 (pp. 152165), Boston, USA.
Heinemann, P. H., Hughes, R., Morrow, C. T., Sommer, H. J., Beelman,
H. J., & West, P. J. (1994). Grading of mushrooms using a machine
vision system. Transactions of the ASAE, 37(5), 16711677.
Kleynen, O., Leemans, V., & Destain, M. F. (2005). Development of a
multispectral vision system for the detection of defects on apples.
Journal of Food Engineering, 69(1), 4149.
Latz, H. W., & Ernes, D. A. (1978). Selective uorescence detection of
citrus oil. Components separated by high-pressure liquid chromatography. Journal of Chromatography, 166, 189199.
Leemans, V., & Destain, M. F. (2004). A real-time grading method of
apples based on features extracted from defects. Journal of Food
Engineering, 61(1), 8389.
Leemans, V., Magein, H., & Destain, M. F. (1998). Defect segmentation
on golden delicious apples by using color machine vision. Computers
and Electronics in Agriculture, 29(2), 117130.
Leemans, V., Magein, H., & Destain, M. F. (1999). Defect segmentation
on Jonagold apples using colour machine vision and a Bayesian
classication method. Computers and Electronics in Agriculture, 23(1),
4353.
Li, Q., Wang, M., & Gu, W. (2002). Computer vision based system for
apple surface defect detection. Computers and Electronics in Agriculture, 36(2), 215223.
Mehl, P. M., Chao, K., Kim, M., & Chen, Y. R. (2002). Detection of
defects on selected apple cultivars using hiperespectral and multispectral image analysis. Applied Engineering in Agriculture, 18(2), 219226.
Mehl, P. M., Chen, Y. R., Kim, M., & Chan, D. E. (2004). Development
of a hyperspectral imaging technique for the detection of apple surface
defects and contaminations. Journal of Food Engineering, 61(1), 6781.
Miller, B. K., & Delwiche, M. J. (1991). Peach defect detection with
machine vision. Transactions of the ASAE, 34(6), 25882597.
Muir, A. Y., Porteus, R. L., & Wastie, R. L. (1982). Experiments in the
detection of incipient diseases in potato tubers by optical methods.
Journal of Agricultural Engineering Research, 27, 131138.

393

Pearson, T. (1995). Machine vision system for automated detection of

stained pistachio nuts, In: Proceedings of the SPIE. Vol. 2345 (pp. 95
103).
Peirs, A., Lammertyn, J., Nicola, B., & De Baerdemaeker, J. (2000). Nondestructive quality measurements of apples by means of nir-spectroscopy. Acta Horticulturae (ISHS), 517, 435440.
Penman, D. W. (2001). Determination of stem and calyx location on
apples using automatic visual inspection. Computers and Electronics in
Agriculture, 33(1), 718.
Poynton, C. (1995). A guided tour of color space. In: Proceedings of the
advanced television and electronic imaging conference. <http://
www.poynton.com/PDFs/Guided_tour.pdf>. Accessed March 2007.
Poynton, C. (1997). Frequently-asked questions about color. <http://
www.poynton.com/PDFs/ColorFAQ.pdf>. Accessed March 2007.
Rehkugler, G. E., & Throop, J. A. (1986). Apple sorting with machine
vision. Transactions of the ASAE, 29(5), 13881397.
Rodriguez-Otero, J., Hermida, M., & Ceneno, J. (1997). Analysis of dairy
products by near-infrared spectroscopy: A review. Journal of Agricultural and Food Chemistry, 45(8), 28152819.
Ruiz, L. A. (1999). Aplicacion de tecnicas de analisis de imagen y
espectrometra infrarroja para la deteccion de danos en ctricos en la
posrecoleccion. Tesis Doctoral. Valencia, Spain: Universidad Politecnica de Valencia.
Ruiz, L. A., Molto, E., Juste, F., Pla, F., & Valiente, R. (1996). Location
and characterization of stem-calyx area on oranges by computer
vision. Journal of Agricultural Engineering Research, 64, 165172.
Schulz, H., Drews, H. H., Quilitzsch, R., & Kruger, H. (1998). Application
of near-infrared spectroscopy for the quantication of quality parameters in selected vegetables and essential oil plants. Journal of Nearinfrared Spectroscopy, 6, 125130.
Shearer, S. A., & Payne, F. A. (1990). Color and defect sorting of bell
peppers using machine vision. Transactions of the ASAE, 33(6),
20452050.
Singh, N., & Delwiche, M. J. (1994). Machine vision methods for defect
sorting stonefruit. Transactions of the ASAE, 37(6), 19891997.
Skjervold, P. O., Taylor, R. G., Wold, J. P., Berge, P., Abouelkaram, S.,
Culioli, J., et al. (2003). Development of intrinsic uorescent multispectral imagery specic for fat, connective tissue, and myobers in
meat. Journal of Food Science, 68(4), 11611169.
Stat Soft, Inc. (2003). Discriminant function analysis. <http://www.statsoft.com/textbook/stdiscan.html>. Accessed March 2007.
Steinmetz, V., Roger, J. M., Molto, E., & Blasco, J. (1999). On-line fusion
color camera and spectrophotometer for sugar content prediction of
apples. Journal Agricultural Engineering Research, 73, 207216.
Vargas, A. M., Kim, M. S., Tao, Y., Lefcourt, A., & Chen, Y. (2004).
Safety inspection of cantaloupes and strawberries using multispectral
uorescence imaging techniques. In: ASAE annual meeting 2004,
ASAE Paper No. 043056, Ottawa, Canada.
Wen, Z., & Tao, Y. (1999). Building rule-based machine vision system for
defect inspection on apple sorting and packing lines. Expert Systems
with Applications, 16, 307313.
Wulfsohn, D., Sarig, Y., & Algazi, R. V. (1993). Defect sorting of dry
dates by image analysis. Canadian Agricultural Engineering, 35(2),
133139.
Yang, C. C., Chao, K., & Chen, Y. R. (2005). Development of
multispectral imaging processing algorithms for food safety inspection on poultry carcasses. Journal of Food Engineering., 69(2),
225234.
Zude, M. (2003). Non-destructive prediction of banana fruit quality using
VIS/NIR spectroscopy. Fruits, 58, 135142.