Computers in Industry 99 (2018) 9–16

Contents lists available at ScienceDirect

Computers in Industry
journal homepage: www.elsevier.com/locate/compind

Detection of citrus fruit and tree trunks in natural environments using a

multi-elliptical boundary model
Tian-Hu Liua,* , Reza Ehsanib , Arash Toudeshkib , Xiang-Jun Zoua , Hong-Jun Wanga
a
College of Engineering, South China Agricultural University, 483 Wushan Road, Guangzhou, Guangdong 510642, China
b
School of Engineering, University of California, Merced, CA 95343, USA

A R T I C L E I N F O A B S T R A C T

Article history:
Received 31 August 2017 Intelligent detection is a key technology in precision agriculture. As items of different color cluster in
Received in revised form 9 March 2018 different non-overlapping elliptical regions, this study proposed a method for constructing a multi-
Accepted 15 March 2018 elliptical boundary model in Cr-Cb co-ordinates to detect citrus fruit and tree trunks in natural light
Available online xxx environments. Here, the detected citrus variety was spring sweet tangerine, and the parameters of the
elliptical boundary models for detecting these fruit and tree trunks solved by color-space transformation
Keywords: and ellipse fitting. A series of image detection experiments were performed to evaluate the method’s
Computer vision performance. The experimental results showed that the correct and false positive percentages in fruit
Natural environments
identification from images were 90.8 and 11.2%, respectively. The number of correctly detected images in
Elliptical boundary model
distinguishing tree trunks from background was 44 of 50 images.
Citrus detection
© 2018 Elsevier B.V. All rights reserved.

1. Introduction identification from the heterogeneous dataset by grouping vertical

A key technology in precision agriculture is intelligent
detection, which is used in target location and robot navigation.
Identifying farmed objects in natural environments is important
for the proper functioning of intelligent agricultural robots.
Currently, machine vision detection is a topic in agricultural
engineering, with various studies having been conducted to detect
fruit, branches and trunks of fruit trees using machine vision. For
example, a computer vision-based method has been used in Tabb
[1] to reconstruct three-dimensional (3-D) fruit trees. Under the
assumption that tree trunks are relatively narrow, vertical shapes,
Lu and Rasmussen [2] have proposed a contrast-based method for
tree detection and shape estimation from ground-plane perspec-
tive images. Shao et al. [3] have presented a method for recognizing
tree trunks based on Hough transformations in CIE L*a*b* color
space (International Commission on Illumination). Shao et al. [4]
have used a backpropagation neural network in recognizing trunks
from an image. Juman et al. [5] have achieved detection rates of
44.0, 59.2, 92.25 and 91.7% using support vector machines (SVM),
neural network, Viola and Jones and their Viola and Jones+ pre-
processing methods for detecting palm tree trunks on static
images. Yıldız [6] has achieved an 88% detection rate in trunk
* Corresponding author.
E-mail address: liuparalake@scau.edu.cn (T.-H. Liu).
edges into potential tree trunks.
Many studies have used machine vision techniques for fruit
identification. Stajnko et al. [7] have estimated the number and
diameter of apple fruit in an orchard during the growing season
using thermal imaging. Zhang and Zhou [8] have studied a three-
layer BP neural network method for segmenting strawberry
images. Miao et al. [9] have presented a watershed algorithm
based on Zernike-moment edge detection to extract grape fruit
contour features. An improved R-G image recognition method has
been proposed by Zhao et al. [10] to recognize apples at night,
which did not account for sheltered and adhesive apples, but the
identification rate in 60 images acquired at night reached 83.7%.
Many studies have reported citrus fruit detection in images. Hue
and saturation thresholds have been applied by Annamalai and Lee
[11] to detect citrus fruit in images. Xu et al. [12] and Zhang et al.
[13] have identified citrus fruit in the tree canopy based on color-
image processing. Hannan et al. [14] have produced an algorithm
using a red chromaticity coefficient to identify oranges and
achieved a 90% detection accuracy with a 4% false-positive rate.
Wang et al. [15] have presented an image processing method for
identifying citrus fruit of different maturities in complex natural
scenes, with an accuracy rate of 92%. Okamoto and Lee [16] have
developed a hyperspectral image processing method to detect
green citrus fruit in individual trees. Lu and Nong [17] have
detected citrus fruit within the tree canopy under natural
illumination conditions, by analyzing the salient edges of

https://doi.org/10.1016/j.compind.2018.03.007
0166-3615/© 2018 Elsevier B.V. All rights reserved.
10 T.-H. Liu et al. / Computers in Industry 99 (2018) 9–16

chromatic aberration maps of R and B channels in the RGB color

model. Kane and Lee [18] have used a monochromatic near-
infrared camera equipped with interchangeable optical band pass
filters to capture citrus fruit images and apply indices and
morphological image processing techniques in segmenting the
images, they achieved an average correct pixel identification of
84.5%. Cubero et al. [19] have reviewed recent works that use color
and non-standard computer vision systems for the automated
inspection of citrus. Kurtulmus et al. [20] have used color, circular
Gabor texture analysis, and a novel ‘eigenfruit’ approach for green
citrus detection, and successfully detected 75.3% of the actual fruit
in the images. Zhao et al. [21] have detected immature green citrus
based on color feature and sum of absolute transformed difference
(SATD) using color images in the citrus grove, and achieved more
than 83% recognition accuracy.
Due to the complex nature of fruit images, no existing algorithm
is totally effective for fruit identification from fruit images. This
study examined a machine vision algorithm suitable for detecting
citrus fruit and tree trunks. Fruit and tree trunk detection are both
important for intelligent fruit-harvesting machines, such as
intelligent canopy shaking machines [22] and fruit picking robots
[23], in which tree trunk detection method is used in robot
navigation and fruit detection method used for choosing target
locations. The present study focused on an image segmentation
algorithm, which was, to the authors’ knowledge, the first study to
date that has examined this strategy.
2. Materials and methods
2.1. Algorithm description
This image segmentation algorithm was similar to a single
supervised-learning method, which was based on artificial
classification and whose computational complexity was much
simpler than algorithms based on machine and deep learning.
This approach generalized well with relatively less training data
and needed much less data storage. A mapping model that can
distinguish multiple regions in an image, should be able to
demonstrate some relationships between a series of sets EJ ¼
fe1 ; e2 ; . . . ; em g and a series of image regionsAJ ¼ fa1 ; a2 ; . . . ; am g
(Fig. 1). If an image can be described by a collection of several non-
overlapping elliptical regions (Fig. 2), then there exists a function G
that can map a region ai of an image into a feature model g = Rk,
with gðai Þ reflecting some of the region’s characteristics.
Y0CbCr color space separates color out into a luma signal (Y0 ) and
two chroma components (Cb and Cr). Experiments here showed
that, if RGB images were transformed into Y'CbCr images, the color
projected in the Cr-Cb co-ordinates for a series of objects a1, a2,
. . . , am were clustered in a series of non-overlapping elliptical
regions e1, e2, . . . , em. Based on this feature, this study proposed a
multi-elliptical boundary model for detecting citrus fruit and tree
Fig. 2. Several elliptical regions that are not overlapping.
trunks. The method included two processes. The first process
calculated the elliptical boundary regions in the Cr-Cb co-
ordinates, which included four steps:
Step 1: Obtain template images for object 1, object 2, . . . ,
object m;
Step 2: Conversion of template images from RGB space to Y0CbCr
space;
Step 3: Sampling skin areas for object 1, object 2, . . . , object m
from template images; and
Step 4: Obtain color clustered regions e1, e2, . . . , em for object 1,
object 2, . . . , object m in the Cr-Cb co-ordinates and construct
multi-elliptical boundary model.
The second process captured images in the natural light
illumination environment and used the multi-elliptical boundary
model to detect fruit and tree trunks. This process had three steps:
Step 1: Capture RGB images;
Step 2: Conversion of captured images from RGB space to Y0CbCr
space; and
Step 3: Segment object 1, object 2, . . . , object m from the
captured images using the multi-elliptical boundary model. The
flow chart of this method was shown in Fig. 3.
2.2. Image acquisition
The 300 images were acquired from a citrus grove (As shown
in Fig. 4.) in Huizhou, Guangdong, China, in April 2017, using a
consumer grade regular digital camera (Sony Cyber-shot DSC-
H50), while the flash light was kept always off. The citrus variety
was spring sweet tangerine. Those images were taken at 1 to 2 m
of distance between the objects (fruit or tree trunks) and the
camera between 10:00 am and 4:00 p.m. Those images were

Fig. 1. Multi-object mapping and segmentation.

 T.-H. Liu et al. / Computers in Industry 99 (2018) 9–16 11

Fig. 3. The processes and flow chat of the algorithm.

Fig. 4. Citrus grove for image acquisition.

saved as 24-bit color JPG format and resized to as 1280  960 2.4. Construct elliptical boundary model
pixels. Microsoft Visual C + + 2013 and Open Source Computer
Vision Library (openCV2.4.9, Intel Corporation) were used in The performance of the elliptical boundary model was
developing software to realize the proposed detection algorithm comparable to that of Gaussian mixture models, and yet, its
on an Intel(R) Core(TM) i7-4600 CPU @ 2.10 GHz 2.69 GHz, 4.00GB computational complexity was as simple as training a single
RAM laptop. Gaussian model [24]. Hsu et al. [25] have concluded that skin color
was distributed as an ellipse in Cr-Cb co-ordinates. The fruit pixels
2.3. Convert images from RGB color space to Y0 CbCr color space colors clustered in the Cr-Cb co-ordinates were within a region
similar to an ellipse (Fig. 5a).
RGB space is the most common color space. However, high In this subsection, the construction of elliptical boundary model
correlation exists between R, G and B colors. After images are was discussed. The moment of order (p and q), with p  0 and q  0,
converted to Y0CbCr color space from RGB color space, the resultant was defined as
signals range from 16 to 235 for Y0 and from 16 to 240 for Cb and Cr. ðð X
q q
Y0CbCr color space is often used in segmenting objects that are mp;q ¼ Crp Cb dCrdCb ¼ Crp Cb ð2Þ
difficult to separate in RGB color space. Here, RGB image was ðCr;CbÞ2R
ðCr;CbÞ2R
converted to Y0CbCr image using Eq. (1)
If p + q  1, the normalized moments of mp,q were given by
Y0 65:481 128:553 24:966 R 16
½ Cr  ¼ ½ 37:797 74:203 112:000 ½ G  þ ½ 128  ð1Þ 1 X q
np;q ¼ Crp Cb ð3Þ
Cb 112:0 93:786 128:214 B 128 aðCr;CbÞ2R

Fig. 5. (a) Color region in Cr-Cb co-ordinates for a fruit skin, and (b) the parameters for an ellipse.
12 T.-H. Liu et al. / Computers in Industry 99 (2018) 9–16

Fig. 6. (a) Template image, (b) Sampled skin, and (c) Color region of skin in Cr-Cb co-ordinates for fruit.

Fig. 7. (a) Template image, (b) Sampled skin, and (c) Color region of skin in Cr-Cb co-ordinates for tree trunk.

Fig. 8. The Fitted elliptical boundaries and their parameters.

where a represents the region’s area. to define the region’s orientation and extent. According to Steger
If p + q  2, the moments relative to the region’s center of et al. [26], the major and the minor axes r1 and r2, as well as the
gravity, which is defined as the central moments, were given by angle u of the ellipse with respect to the column axis (Fig. 5b) were
given by
1 X
mp;q ¼ ðCr  n1;0 Þp ðCb  n0;1 Þq ð4Þ 8 rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
aðCr;CbÞ2R >
>
> r ¼ 2ðm2;0 þ m0;2 þ ðm2;0  m0;2 Þ2 þ 4m21;1 Þ
>
> 1
>
> rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
The co-ordinates of the region’s center were given by <
8 X X r2 ¼ 2ðm2;0 þ m0;2  ðm2;0  m0;2 Þ2 þ 4m21;1 Þ ð6Þ
>
>
>
> Cr Cb >
>
< m1;0 ðCr;CbÞ2R m0;1 ðCr;CbÞ2R >
>
> 1 2m1;1
Cr ¼ ¼ X Cb ¼ ¼ X ð5Þ : u ¼ arctan
>
> m0;0 1 m0;0 1 2 m2;0  m0;2
:
ðCr;CbÞ2R ðCr;CbÞ2R
A pixel was classified as a fruit pixel if its chrominance
Assuming a region’s moments of order 1 and 2 were obtained vectors (Cr and Cb) were inside the fitted ellipse region of fruit
from an ellipse, the second central moments (p + q  2) were used skin, or it was classified as a tree trunk pixel if its chrominance
 T.-H. Liu et al. / Computers in Industry 99 (2018) 9–16 13

Fig. 9. Fruit detection results from (a) a thin density image and (b) a thick density image.

Fig. 10. Fruit numbers counted by human, the correct and the finally counted by computer along with the false positive and the missed.

Fig. 11. The performance of the algorithm in detecting each of those 50 images.

vectors were inside the fitted ellipse region of tree trunk skin. MATLAB 2016a software (Mathworks, Inc., Natick, MA, USA)
Otherwise, it was classified as “other” pixels. The criteria was was used for designing program calculating parameters for these
expressed as elliptical boundary models. Template images and sample skins for
8 calculating the elliptical boundary model of fruit and tree trunks,
> Fruit pixel; if vectors ðCr; Cb Þare inside the ellipse region of fruit skin .
< along with regions of color clustered in Cr-Cb co-ordinates for
Tree trunk pixel; if vectors ðCr; Cb Þare inside the ellipse region of tree trunk skin .
>
: those two sampled skins are shown in Figs. 6 and 7, respectively.
Other; otherwise:
The fitted elliptical boundary models and their parameters are
ð7Þ shown in Fig. 8.
14 T.-H. Liu et al. / Computers in Industry 99 (2018) 9–16

Table 1 Table 2
Fruit detection accuracy. Sampled fruit skins along with their lightness and their fitted elliptical boundary
models.
Items Results Standard deviation
Images 50 – Fruit skin Lightness Fitted elliptical model
Total fruit number 1299 – 34%
CDP 90.8% 0.044
FPP 11.2% 0.053
MP 9.2% 0.044

51%

3. Results and discussion

To demonstrate how the proposed method works, the

59%
intermediate results for three examples, which were randomly
chosen from those captured images, were displayed in the
following subsections.

3.1. Fruit detection 67%

Noise was masked by treating images with a digital-morpho-

logical-opening operation and area-thresholding after segmenta-
tion using the elliptical boundary model. The digital- 72%
morphological-opening operation was a geometry-based filter,
which removed isolated noises, points, and burrs. The detected
final results for two images are shown in Fig. 9, in which the
identified fruit labeled with red circles. The algorithm was not able 77%
to detect all fruit in the images, and there were detections in parts
of the image that were not fruit, being false positives. The false
positives were mostly a consequence of foliage, branches and stem
coverage as well as irregular tree canopy shading that split a fruit’s 92%
contour into several parts. Because the color region for citrus
flowers clustered in the Cr-Cb co-ordinates overlapped with that of
the fruit, some citrus flowers were wrongly detected as fruit, while
a few fruit were undetected. The missed fruit were mostly a
consequence of shelter, uneven lighting conditions, fruit color
differences, irregular tree canopies, foliage coverage, fruit adhered statistical significance than randomly arranged single values. The
together, distance between the camera and tree canopy, and averages of the CDP, FPP, and MP and their standard deviations
variable fruit sizes. The algorithm was applied to 50 images as a showed that the correct fruit identification accuracy was 90.8%,
validating dataset (Fig. 10), and fruit quantities determined both by with the false-positive and missed percentages at 11.2 and 9.2%,
manual and algorithm counting. The false-positive and missed respectively (Table 1). The standard deviation measured the
fruit quantities were determined by manual and comparison. volatility of the results, and, the larger the standard deviation, the
To evaluate the performance of the proposed algorithm in greater the results’ volatility. Here, the standard deviation values
detecting fruit, three performance metrics were defined. The first indicated that FPP volatilities among different images were large.
metric was the correct detection percentage (CDP) and expressed Correctly:detected:f ruit:number
in Eq. (8). The second metric was the false-positive percentage CDP ¼  100% ð8Þ
The:total:f ruit:number:in:an:image
(FPP) and expressed in Eq. (9). The third metric was the missed
percentage (MP) and expressed in Eq. (10). The algorithm’s
performance in detecting the 50 images indicated that the CDP, False:positive:number
FPP, and MP were quite different among different images (Fig. 11). FPP ¼  100% ð9Þ
The:total:f ruit:number:in:an:image
The average values and their standard deviation possessed more

Fig. 12. (a) The original tree trunk image, and (b) resulting image after segmentation.
 T.-H. Liu et al. / Computers in Industry 99 (2018) 9–16 15

Table 3 The number of correctly detected images in distinguishing trunks

Sampled tree trunk skins along with their lightness and their fitted elliptical
from the background was 44 of 50 images. The direction of solar
boundary models.
illumination was found to affect detection quality, with detected
Trunk skin Lightness Fitted elliptical model results from images taken in the direction of the sun having good
6% quality with little noise. However, images taken in the opposite
direction possessed much noise.
In terms of the detection rate, the present method was
greater than SVM and neural network-based methods reported
17% by Juman et al. [5] and almost the same as reported by Yıldız [6],
in which he correctly detected 88% of tree trunks in a
heterogeneous dataset. The disadvantage of SVM and neural
networks is that the training procedure requires a lot of time and
29% training with samples. Overfitting often occurs when the
training dataset contains some incorrect samples in the BP
neural network.

51% 3.3. Discussion

The obvious advantage of the present method was that its

63%
95%
Missed:f ruit:number
MP ¼  100%
The:total:f ruit:number:in:an:image
The accuracy of this algorithm was a little greater than that
reported by Hannan et al. [14] and Kane and Lee [18], whose
accuracy was 90 and 84.5%, respectively. However, the present
accuracy was lower than the 92% accuracy reported by Wang et al.
[15]. The FPP was also greater than that reported by Hannan et al.
[14] using their method. The average summed identification time
for a 1280  960 pixel image was 0.328 s.
3.2. Trunk detection
In an example of citrus tree trunk detection, the results
indicated that the algorithm masked most background materials,
such as soil and weeds (Fig. 12). However, some noise still existed.
computational complexity was as simple as training a single
Gaussian model. However, lighting changed the position and
shape of elliptical regions, which led to missed and false-
positive detections. Fitted elliptical boundary models under
different lighting for citrus fruit and tree trunk skins are shown
in Tables 2 and 3. The L component of HSL (Hue-Saturation-
Lightness) referred to the lightness of color which ranged from 0
to 100%, the smaller the value, the darker the color, and the
closer it is to black. In contrast, the bigger the value, the brighter
the color, and the closer it is to white. As the lightness here was
between 30 and 70%, the position and shape of elliptical
boundary models changed very little. But, as lightness varied
from 0 to 30% and from 70 to 100%, position and shape clearly
change. Although different lighting environments affected the
ð10Þ accuracy and robustness of subsequent image detection, if the
brightness of image was corrected using brightness correction
methods, the position of elliptical boundary model changed and
its area became larger than the original. A template image and
sampled fruit skin along with the region of color, clustered in the
Cr-Cb co-ordinates after brightness correction using the Retinex
algorithm [27] is shown in Fig. 13; the color region area in
Fig. 13c was much larger than that in Fig. 6c. If the area of
elliptical boundary model became larger, the false-positive rate
increased. Therefore, different elliptical models should be used
to segment the same object under different lighting conditions.
In the future, the algorithm will be improved by taking the
lighting parameter into account, such that the parameters of
elliptical models would be automatically adjusted according to
the environment’s lightness.

Fig. 13. (a) Template image, (b) Sampled skin, and (c) Color region in Cr-Cb co-ordinates after brightness correction.
16 T.-H. Liu et al. / Computers in Industry 99 (2018) 9–16
4. Conclusion
This study proposed a method for constructing a multi-elliptical
boundary model in Cr-Cb co-ordinates to detect citrus fruit and tree
trunks in natural light environments. In the solution proposed here,
the construction of an elliptical boundary model and the related
detection process were discussed. Before detection, images were
converted from RGB color space to Y0CbCr color space. The algorithm
was not able to detect all fruit in the images, with the correct
detection percentage and false positives in fruit detection at 90.8 and
11.2%, respectively. The number of correctly detected images in
distinguishing tree trunks from background was 44 of 50 images, and
the direction of solar illumination affected detection results.
Lightness changed the position and shape of ellipse models, and
led to missed and false-positive detections. However, if the image
brightness was corrected using brightness correction methods, the
position of elliptical boundary model changed and its area became
larger than the original, which led to an increased false-positive
rate. Therefore, under different lighting conditions, different
elliptical models should be used to segment the same object.
Acknowledgements
This research is co-supported by Guangdong Science and
Technology Plan Project with Research Grant 2017A010102024,
National Key R & D Plan of China with Research Grant
2017YFD0700100 and National Science Project of China with
Research Grant 31571568.
References
[1] A. Tabb, Three-dimensional Reconstruction of Fruit Trees by a Shape from
Silhouette Method, ASABE paper no. 096138. St. Joseph, Mich, ASABE, 2009.
[2] Y. Lu, C. Rasmussen, Tree trunk detection using contrast templates, 18th IEEE
International Conference on Image Processing (2011) 1253–1256.
[3] L. Shao, X. Chen, B. Milne, P. Guo, A novel tree trunk recognition approach for
forestry harvesting robot, 9th IEEE Conference on Industrial Electronics and
Applications (2014) 862–866.
[4] Y. L. Shao, J. Mu, G. Liu, H. Dong, P. Guo Liu, The trunk of the image recognition
based on BP neural network, 2014 IEEE International Conference on
Mechatronics and Automation (2014) 1800–1805.
[5] M.A. Juman, Y.W. Wong, R.K. Rajkumar, L.J. Goh, A novel tree trunk detection method
for oil-palm plantation navigation, Comput. Electron. Agric. 128 (2016) 172–180.
[6] T. Yıldız, Detection of Tree Trunks as Visual Landmarks in Outdoor
Environments, Bilkent University, 2010.
[7] D. Stajnko, M. Lakota, M. Hocevar, Estimation of number and diameter of apple
fruits in an orchard during the growing season by thermal imaging, Comput.
Electron. Agric. 42 (2004) 31–42.
[8] T.Z. Zhang, T.J. Zhou, Strawberry harvesting robot: I. Segmentation of
strawberry image by BP neural network, J. Chin. Agric. Univ. 9 (4) (2004)
65–68.
[9] Y.B. Miao, Z.M. Wang, Q. Liu, Application of Zernike-moment-based
watershed segmentation on fruit features extraction, Trans. CSAE 29 (1)
(2013) 158–163.
[10] D.A. Zhao, X.Y. Liu, Y. Chen, W. Ji, W.K. Jia, C.L. Hu, Image recognition at night for
apple picking robot, Trans. Chin. Soc. Agric. Mach. 26 (3) (2015) 15–21.
[11] P. Annamalai, W.S. Lee, Citrus Yield Mapping System Using Machine Vision,
ASAE Paper No. 031002. St. Joseph, ASAE, 2003.
[12] H.R. Xu, Z.Z. Ye, Y.B. Ying, Identification of citrus fruit in a tree canopy using
color information, Trans. CSAE 21 (5) (2005) 98–101.
[13] Y.J. Zhang, L. Deng, M.Z. Li, Estimation of citrus yield based on image
processing, Trans. Chin. Soc. Agric. Mach. 40 (9) (2009) 97–99.
[14] M. Hannan, T.F. Burks, D.M. Bulanon, A machine vision algorithm for orange
fruit detection, The CIGR Ejournal. Manuscript (2009)1281. Vol XI.
[15] J.H. Wang, Q. Ouyang, Q.S. Chen, Adaptive recognition of different maturity
citrus in natural scene, Opt. Optoelectron. Technol. 7 (5) (2009) 56–58 62.
[16] H. Okamoto, W.S. Lee, Green citrus detection using hyperspectral imaging,
Comput. Electron. Agric. 66 (2009) 201–208.
[17] J. Lu, S. Nong, Detecting citrus fruits and occlusion recovery under natural
illumination conditions, Comput. Electron. Agric. 110 (2015) 121–130.
[18] K.E. Kane, W.S. Lee, Multispectral Imaging for In-field Green Citrus
Identification, Asabe Paper No. 073025. St. Joseph, ASABE, 2007.
[19] S. Cubero, W.S. Lee, N. Aleixos, F. Albert, J. Blasco, Automated systems based on
machine vision for inspecting citrus fruits from the field to postharvest—a
review, Food Bioprocess Technol. 9 (10) (2016) 1623–1639.
[20] F. Kurtulmus, W.S. Lee, A. Vardar, Green citrus detection using ‘eigenfruit’,
color and circular Gabor texture features under natural outdoor conditions,
Comput. Electron. Agric. 78 (2) (2011) 140–149.
[21] C. Zhao, W.S. Lee, D. He, Immature green citrus detection based on colour
feature and sum of absolute transformed difference (SATD) using colour
images in the citrus grove, Comput. Electron. Agric. 124 (2016) 243–253.
[22] T.H. Liu, R. Ehsani, A. Toudeshki, X.J. Zou, H.J. Wang, Experimental study of
vibrational acceleration spread and comparison using three citrus canopy
shaker shaking tines, Shock Vib. 2017 (1) (2017) 1–9.
[23] T.H. Liu, Z.H. Liang, J.D. Guo, Experimental comparison of litchi fruit stalk
cutting force, Appl. Eng. Agric. 28 (2) (2012) 297–302.
[24] J.Y. Lee, S.I. Yoo, An elliptical boundary model for skin color detection,
Proceedings of the International Conference on Imaging Science, Syst.
Technol., CSREA Press, Las Vegas, Nevada, 2002, pp. 579–584.
[25] R.L. Hsu, M. Abdel-Mottaleb, A.K. Jain, Face detection in color images, IEEE
Trans. Pattern Anal. Mach. Intell. 24 (5) (2002) 696–706.
[26] C. Steger, M. Ulrich, C. Wiedemann, Machine Vision Algorithms and
Applications, Wiley, Weinheim, 2008.
[27] N.W. Daw, Goldfish retina organization for simultaneous colour contrast,
Science 158 (3803) (1967) 942–944.