Deep Feature-Based Classifiers for Fruit Fly Identification (Diptera: Tephritidae)

Full text

Deep Feature-based Classifiers for Fruit Fly
Identification (Diptera: Tephritidae)
Matheus M. Leonardo∗, Tiago J. Carvalho†, Edmar Rezende‡, Roberto Zucchi§, Fabio A. Faria∗,
∗Institute of Science and Technology - Universidade Federal de S ˜
ao Paulo (UNIFESP), S˜
ao Jos´
e dos Campos, Brazil
Email: matheus.macedo.leonardo@gmail.com and ffaria@unifesp.br
†Federal Institute of S˜
ao Paulo, Campinas, Brazil
Email: tiagojc@ifsp.edu.br
‡University of Campinas, Campinas, Brazil
Email: edmar.rezende@gmail.com
§Luiz de Queiroz College of Agriculture, University of Sao Paulo, Piracicaba, Brazil
Email: rzucchi@usp.br
Abstract—Fruit flies has a big biological and economic im-
portance for the farming of different tropical and subtropical
countries in the World. Specifically in Brazil, third largest fruit
producer in the world, the direct and indirect losses caused
by fruit flies can exceed USD 120 million/year. These losses
are related to production, the cost of pest control and export
markets. One of the most economically important fruit flies
in the America belong to the genus Anastrepha, which has
approximately 300 known species, of which 120 are recorded in
Brazil. However, less than 10 species are economically important
and are considered pests of quarantine significance by regulatory
agencies. The extreme similarity among the species of the genus
Anastrepha makes its manual taxonomic classification a nontrivial
task, causing onerous and very subjective results. In this work,
we propose an approach based on deep learning to assist the
scarce specialists, reducing the time of analysis, subjectivity of
the classifications and consequently, the economic losses related
to these agricultural pests. In our experiments, five deep features
and nine machine learning techniques have been studied for the
target task. Furthermore, the proposed approach have achieved
similar effectiveness results to state-of-art approaches.
I. INTRODUCTION
Anastrepha is the most diverse genus in the Americas, with
over 300 species known in the tropics and subtropics regions.
In Brazil, 120 species are recorded [1], but less than 10 species
are considered as agricultural pests.
Species of Anastrepha are known as fruit flies, as females
lay their eggs in healthy fruit, and larvae feed inside the fruit.
Later on, larvae leave the fruit and pupate in the soil and, after
some days, the adults emerge.
Due to the damage caused by larvae in commercialized
fruits, some species of Anastrepha are of significant economic
importance. Furthermore, some Anastrepha species are also
economically important as quarantine pests as they hinder
the international trade of fruits. Consequently, an accurate
identification of quarantine pests of fruits for exportation is
highly relevant, due to the strict trade quarantines to prevent
their spread.
However, identification of cryptic species (closely related
and morphologically similar) is problematic. Some species
of fruit fly quarantine pests comprises complexes of cryptic
species, and it is difficult to delimit the species boundaries
within these complexes.
Identification of cryptic species of agricultural importance is
a huge challenge, especially those with a quarantine status. On
the other hand, misidentification are problematic for quarantine
restrictions, control programs (e.g. integrated pest manage-
ment) and basic studies (biology, geographical distribution,
plant hosts, and natural enemies).
Anastrepha fraterculus, the South America Fruit Fly, is
a complex of species in the Americas. It is a major pest
only in some areas of its geographical distribution from
Mexico to northern Argentina. Another complex with species
economically important is the Anastrepha obliqua complex.
Species of these two complexes are widely distributed in Brazil
and the real challenge is to find out which species of these
complexes are actually agricultural pests.
Several techniques (e.g. crosses, morphometric and molec-
ular analyses [2], [3]) have been used for clarifying the
identity of the species within these complexes. Recently, image
analysis techniques were used to identifying three species of
Anastrepha, two of them of quarantine importance ( [4]–[6]).
In insect identification literature, more specifically in iden-
tification of fruit flies of the genus Anastrepha, different
methods rises in the last years [2], [3]. Martineau et al. [7]
compiled the most recent works, an overall forty four works,
based on image processing and machine learning techniques
in a survey.
Different literature works associating image processing and
machine learning have been proposed to solve problems that
involves wild life. Digital Automated Identification SYstem
(DAISY) classifies spiders, pollen grain, and butterfly through
a semi-automated classification approach based on principal
component analysis (PCA) [8]. SPecies IDentified Automati-
cally (SPIDA-web) identifies Australian spiders to distinguish
121 species using Daubechies 4wavelet function [9]. Auto-
matic Bee Identification System (ABIS) recognizes bee species
of genus Bombus,Colletes, and Andrena through the use of
support vector machine (SVM) and kernel discriminate anal-
ysis techniques [10], [11]. DAIIS tool classifies a sample of

120 owlflies (Neuroptera: Ascalaphidae) based on their wing
outlines using Elliptic Fourier coefficients and SVMs [12],
[13]. In [14], an insect recognition system that combines
different visual properties such as color, texture, shape, scale-
invariant feature transform (SIFT) [15] and histogram of ori-
ented gradients (HOG) [16] features has been created through
the use of multiple-task sparse representation and multiple-
kernel learning (MKL) techniques [17]. However, a very im-
portant fact needs to be pointed out as motivation of this work.
Despite the biological and economic importance of the Tephri-
tidae family (fruit fly), only three papers have been found in
the literature. The first one adopted a successful framework
of classifier selection and fusion [18], which combines many
global image descriptors and machine learning techniques for a
multimodal classification approach, using image of wings and
aculei of three species of the fraterculus group: A. fraterculus
(Wied.), A. obliqua (Macquart) and A. sororcula Zucchi [4].
In the second work, a mid-level representation based on
BossaNova approach [19] has been adopted to improve the
effectiveness results achieved in previous experiments with
global image descriptors [6]. Finally, in the third work, the
authors proposed a sparse representation using SIFT features
densely sampled as input for two machine learning techniques
(multi-layer Max-pooling ScSPM [20] and linear SVM [10]).
It has performed experiments with three unreported genus and
twenty species [21].
In this work, we take advantage of well known deep learning
architectures, designed to object recognition problems, to
extract relevant features for fruit fly identification. Associating
these features with simple classifiers, we are able to achieve an
effective accuracy on fruit fly classification without necessity
of laborious hand-crafted features extraction. Furthermore,
we compare the effectiveness of our results with different
machine learning techniques using those hand-craft features
to support the development of a real-time system for fruit fly
identification of the genus Anastrepha. We believe that this
system can be a good solution for more precise identification,
allowing time reduction, costs in performing and assisting the
few specialists in their tasks.
The main contributions of this work are:
•Decrease of complexity in features engineering process
when compared with state-of-the-art methods;
•Proposition of a new approach for fruit fly identification
task of the genus Anastrepha based on a deep Convolu-
tional Neural Networks (CNN) model combined with a
transfer learning approach;
•An effectiveness analysis among the best deep learning
architecture and the state-of-art approaches existing in the
literature;
The remainder of this paper is organized as follows. Sec-
tion II describes different feature extraction techniques used in
this works for insect identification task. Section III shows the
experimental protocol we devised to validate the work while
Section IV discusses the results. Finally, Section V concludes
the paper and points out future research directions.
II. MATE RIALS A ND ME THODS
The typical image classification pipeline is composed of the
three following steps: (i) local visual feature extraction, which
extracts information directly from the image pixels, (ii) mid-
level feature extraction, which makes the representation more
general, aggregating abstraction to the model, and (iii) super-
vised classification, a machine learning technique allowing the
extraction of a general model from the data.
Usually, stages (i) and (ii) are strongly depend from a ex-
pensive hand-craft features extraction process, where an expert
looks for the best algorithm (or combination of algorithms)
that better represent the problem samples.
We take advantage of robust Deep Neural Networks (DNN)
architectures, to extract a set of features which are very specific
for the problem, avoiding the laborious hand-craft features
extraction step.
In a nutshell, our method uses a DNN architecture (with-
out the top layer responsible by classification) to extract
features, named bottleneck features. Bottleneck term refers
to a topology of a neural network where the hidden layer
has significantly lower dimensionality than the input layer,
assuming that such layer — referred to as the bottleneck —
compresses the information needed for mapping the neural
network input to the neural network output, increasing the
system robustness to noise and overfitting. Conventionally,
bottleneck features are the output generated by the bottleneck
layer [22]. These bottleneck features are used to train a binary
classifier.
A. Deep Learning Architectures
In the proposed approach, we have used deep convolu-
tional neural networks based on VGG (VGG16 and VGG19),
GoogLeNet (Inception V3 and Xception) and ResNet (ResNet-
50) architectures, shown in Figure 1, pre-trained for object
detection task on the ImageNet dataset.
1) VGG Architecture: The VGG networks [23] with 16
layers (VGG16) and with 19 layers (VGG19) were the basis
of the Visual Geometry Group (VGG) submission in the
ImageNet Challenge 2014, where the VGG team secured the
first and the second places in the localization and classification
tracks respectively.
The VGG architecture is structured starting with five blocks
of convolutional layers followed by three fully-connected
layers. Convolutional layers use 3×3kernels with a stride of
1 and padding of 1 to ensure that each activation map retains
the same spatial dimensions as the previous layer. A rectified
linear unit (ReLU) activation is performed right after each
convolution and a max pooling operation is used at the end
of each block to reduce the spatial dimension. Max pooling
layers use 2×2kernels with a stride of 2 and no padding to
ensure that each spatial dimension of the activation map from
the previous layer is halved. Two fully-connected layers with
4096 ReLU activated units are then used before the final 1000
fully-connected softmax layer.
A downside of the VGG16 and VGG19 models is that they
are more expensive to evaluate and use a lot of memory and

Fig. 1. VGG16, VGG19, Inception V3, Xception and ResNet-50 architectures.

parameters. VGG16 has approximately 138 million param-
eters and VGG19 has approximately 143 million parameters.
Most of these parameters (approximately 100 million) are in
the first fully-connected layer, and it was since found that these
fully-connected layers could be removed with no performance
downgrade, significantly reducing the number of necessary
parameters.
2) GoogLeNet Architecture: The GoogLeNet architecture
was introduced as GoogLeNet (Inception V1), later refined as
Inception V2 and recently as Inception V3 [24].
While Inception modules are conceptually convolutional
feature extractors, they empirically appear to be capable of
learning richer representations with less parameters. A tradi-
tional convolutional layer attempts to learn filters in a 3D
space, with 2 spatial dimensions (width and height) and a
channel dimension. Thus, a single convolution kernel is tasked
with simultaneously mapping cross-channel correlations and
spatial correlations.
The idea behind the Inception module is to make this
process easier and more efficient by explicitly factoring it into
a series of operations that would independently look at cross-
channel correlations and at spatial correlations.
The Xception architecture [25] is an extension of the
Inception architecture which replaces the standard Inception
modules with depthwise separable convolutions. Instead of
partitioning input data into several compressed chunks, it maps
the spatial correlations for each output channel separately,
and then performs a 1×1depthwise convolution to cap-
ture cross-channel correlation. This is essentially equivalent
to an existing operation known as a “depthwise separable
convolution”, which consists of a depthwise convolution (a
spatial convolution performed independently for each channel)
followed by a pointwise convolution (a 1×1convolution across
channels). We can think of this as looking for correlations
across a 2D space first, followed by looking for correlations
across a 1D space. Intuitively, this 2D + 1D mapping is easier
to learn than a full 3D mapping.
Xception slightly outperforms InceptionV3 on the ImageNet
dataset, and vastly outperforms it on a larger image classifi-
cation dataset with 17,000 classes. Most importantly, it has
a similar number of parameters as Inception V3, implying
a greater computational efficiency. Xception has 22,855,952
trainable parameters while Inception V3 has 23,626,728 train-
able parameters.
3) ResNet Architecture: Residual Networks (ResNets) [26]
are deep convolutional networks where the basic idea is to skip
blocks of convolutional layers by using shortcut connections
to form blocks named residual blocks. These stacked residual
blocks greatly improve training efficiency and largely resolve
the degradation problem present in deep networks.
In ResNet-50 architecture, the basic blocks follow two sim-
ple design rules: (i) for the same output feature map size, the
layers have the same number of filters; and (ii) if the feature
map size is halved, the number of filters is doubled. The down-
sampling is performed directly by convolutional layers that
have a stride of 2 and batch normalization is performed right
after each convolution and before ReLU activation.
When the input and output are of the same dimensions,
the identity shortcut is used. When the dimensions increase,
the projection shortcut is used to match dimensions through
1×1convolutions. In both cases, when the shortcuts go across
feature maps of two sizes, they are performed with a stride of
2. The network ends with a 1,000 fully-connected layer with
softmax activation. The total number of weighted layers is 50,
with 23,534,592 trainable parameters.
B. Transfer Learning
Transfer learning consists in transferring the parameters of
a neural network trained with one dataset and task to another
problem with a different dataset and task [27].
Many deep neural networks trained on natural images
exhibit a curious phenomenon in common: on the first layers
they learn features that appear not to be specific to a particular
dataset or task, but general in that they are applicable to many
datasets and tasks. Features must eventually transition from
general to specific by the last layers of the network. When
the target dataset is significantly smaller than the base dataset,
transfer learning can be a powerful tool to enable training a
large target network without overfitting.
In the proposed approach, we have used VGG16, VGG19,
Inception V3, Xception and ResNet-50 as the base models,
pre-trained for object detection task on the ImageNet dataset.
The ImageNet is a public dataset containing 1.28 million
natural images of 1,000 classes.
III. EXP ERIME NTAL SETTINGS
We now discuss the experimental setup, including the
dataset, details on the image acquisition, and machine learning
techniques used in this work.
A. Dataset
In our experiments, we have used lab-based samples [7]
of the specimens A. fraterculus,A. obliqua and A. sororcula
from the collection of the Instituto Biol´
ogico of S˜
ao Paulo
(e.g., Figures 2(a-c)). Specimens have been collected through
McPhail-type traps (Figure 2b) and reared flies from fruit as
well.
(a) (b) (c)
Fig. 2. (a) A fruit fly example (drawing) [28]; (b) a McPhail-type trap; and
(c) a fruit fly laying eggs. Extracted from [4].
The dataset used in this work is composed of 301 images
(resolution 2560 ×1920) and divided into three different cate-
gories: A. fraterculus (100), A. obliqua (101), and A. sororcula
(100). It consists of pictures of specimens reared from samples

A. fraterculus A. obliqua A. sororcula
Fig. 3. Example of wings of each specie studied. Extracted from [4].
of fruit trees in experimental and commercial orchards in
the state of S˜
ao Paulo, Brazil, stored in the Department of
Entomology and Acarology ESALQ, Piracicaba, SP, Brazil
and in the Biological Institute, Campinas, SP, Brazil.
Figure 3 shows examples of the three species used in this
work.
B. Machine Learning Techniques
We have used nine different machine learning techniques:
Decision Tree (DT), k-Nearest Neighbor (kNN) with k=
{1,3,5,7}, Multiple Layer Perceptron (MLP), Na¨
ıve Bayes
(NB), Stochastic Gradient Descent (SGD) and Support Vector
Machine (SVM) using linear kernel.
The proposed methods have been implemented using Python
3.5, Keras 2.0.31, TensorFlow 1.0.12and Scikit-Learn3. All
performed tests have been executed in a machine with an In-
tel(R) Xeon(R) CPU E5-2620 2.00GHz processor with 96GB
of RAM and two Nvidia Titan Xp GPUs.
IV. RES ULTS AND DISCUSSION
We have performed two experiments with objective to sup-
port a system for fruit fly identification. Firstly, a comparative
study among five different deep features and nine machine
learning techniques has been performed. Next, a comparison
among the best tuple (feature +learning technique) against
two of the best literature [4], [6] approaches. For all effective-
ness experiments, the average accuracies in the 5-fold cross-
validation protocol have been computed (three partitions for
training set, one for validation set, and one for test set).
A. Effectiveness Analysis
In this section, we have performed a comparative study
among five deep features based on deep learning architectures
(Inception, ResNet, VGG16, VGG19, and Xception).
Table I shows effectiveness results for all deep features and
machine learning techniques. We can observe that bottleneck
features extracted using VGG16 architecture has achieved
eight of the best effectiveness results among nine released
learning techniques (in blue). Furthermore, we can observe
that Support Vector Machines (SVM) technique using linear
kernel has achieved the best effectiveness results for all of the
five deep learning architectures applied for features extraction
1https://keras.io
2https://www.tensorflow.org
3http:http://scikit-learn.org/stable/ (As of January, 2018)
(in gray cell) released in this work. In addition, SVM technique
using VGG16 feature was the best tuple (feature +learning
technique) performed in this work with 95.68% of average
accuracy (in blue text and gray cell).
B. The Best Approaches of the literature
The second experiment performed compares the best tu-
ples (deep feature +learning technique) of the previous
experiment (Inception+SVM, ResNet+SVM, VGG16+SVM,
VGG19+SVM, and Xception+SVM) against state-of-the-art
methods (LCH+SVM [4] and F-SIFT+MLP [6]). LCH+SVM
is a support vector machine technique with polynomial kernel
using a generic color image descriptor called Local Color
Histogram [29]. F-SIFT+MLP is a multiple layer perceptron
technique using a Bow-of-Words (BoW) representation based
on BossaNova [19] with keypoint detector FAST [30] and local
feature detector SIFT [15]
Figure 4 shows the effectiveness results among the best
tuples (feature +learning technique) and the best baseline
existing in the literature. Although VGG16+SVM (in blue)
has achieved the best mean accuracy (95.68%), when we
compute the confidence interval with significance level of
0.05, it is possible to observe that there is no statistically
significant difference among five deep learning approaches and
the two baselines (in red) from the literature LCH+SVM [4]
(93.50%) and F-SIFT+MLP [6] (94.67%). However, it is very
important to note that LCH+SVM approach has achieved
good effectiveness results by extracting color properties from
enhanced image (e.g., segmentation and morphological opera-
tions). Therefore, this color-based histogram approach may not
be able to be used in real-time systems, unlike the other two
compared approaches (F-SIFT+MLP and VGG16+SVM).
Moreover, in [6], we could see that the F-SIFT+MLP
approach need to find the best tuple (keypoint detector and
feature extractor), as well as, to perform the BoW process
for finally achieving good results in the fruit fly identification
task. This fact does not occur in our proposed approach based
on deep learning techniques, since we only need to find the
architecture configuration that best describes the data of the
target application.
In relation to the properties extracted by compared ap-
proaches, we can also verify differences between them. LCH
deals with color properties, F-SIFT works with keypoint de-
tector that are gradient dependent (shapes, edges, and corners)
and deep learning works with the combination of different

TABLE I
EFFE CTI VEN ESS R ESU LTS (IN %) AMONG FIVE DEEP LEARNING ARCHITECTURES AND NINE MACHINE LEARNING TECHNIQUES FOR A 5 -FO LD
CRO SS-VAL IDATI ON PROT OCO L. IN B LUE A RE T HE BE ST IM AGE DE EP FE ATUR ES FO R EACH M ACH INE L EAR NIN G TEC HN IQU E. IN G RAY CEL L ARE T HE
BE ST MAC HIN E LEA RNI NG T ECH NIQ UES F OR EAC H DEE P LE ARN ING A RCH ITE CTU RE US ED FO R BOT TLE NEC K FEAT URE S EXT RACT ION .
Deep Features Machine Learning Techniques
DT KNN1 KNN3 KNN5 KNN7 MLP NB SGD SVM
Inception 57.83 71.09 70.09 70.09 66.77 87.7 54.13 88.03 88.37
ResNet 65.43 75.08 77.08 77.74 80.06 89.03 73.40 89.70 90.36
VGG16 75.75 84.39 89.02 87.71 87.73 95.02 75.75 93.36 95.68
VGG19 72.1 84.72 82.08 80.09 81.39 92.68 67.13 91.35 94.34
Xception 51.48 60.79 56.12 55.77 55.44 68.33 51.82 78.41 78.74
properties (e.g., corners, edge/color conjunctions, and tex-
ture [31]). In a real system, if some visual property is deficient,
it is believed that only the deep learning approach might be
able to extract features of the available data and to achieve
some good result.
Finally, we can use all of three approaches (LCH, F-SIFT,
and VGG16) working together and complementary so that the
final effectiveness results become more robust. Another idea
is to mix concepts BoW and feature deep like in paper [32].
70
75
80
85
90
95
100
Xception+SVMLinear
Inception+SVMLinear
Resnet+SVMLinear
LCH+SVMPoly
VGG19+SVMLinear
F−SIFT+MLP
VGG16+SVMLinear
Mean Accuracy (%)
The best learning techniques for each image feature
Fig. 4. Effectiveness results for each image feature with 95% confidence
interval (CI), i.e, a significance level of 0.05. In blue is SVMLinear using
VGG16 feature that has achieved the best mean accuracy.
V. CONCLUSION
In this work, we proposed a method that take advantage of
deep learning architectures associated with transfer learning
approach for fruit fly identification task. Extracting bottleneck
features, we characterize images of three species of the genus
Anastrepha. Then we use a shallow classifier to detect the
correct class of the image.
Five well known deep architectures and nine machine
learning techniques have been compared, achieving a better
accuracy of 95.68% when associating VGG16+SVM.
When compared against state-of-art approaches, our method
perform better in evaluated images, without any additional
image processing enhancement operation. This fact is very im-
portant for constructing a real-time system that helps fighting
back pests in agriculture.
Unlike the mid-level representation approach (F-SIFT) that
need to choose the best combination keypoint detector and
feature extractor, adding the BoW process.
The use of deep learning techniques applied to fruit fly iden-
tification with transfer learning approach is quite new. Mainly
because of the lack of samples to train a deep architecture from
scratch, this approach brings a feasible way to use CNNs in the
target task, eliminating the necessity of generate hand crafted
features.
As future work, we intend to perform experiments with
species and learning techniques as classifier ensemble. Other
work might be the development of a mobile system to assist
the few experts from the biology area on their field works.
VI. ACK NOW LE DGMEN T
The author thanks the support of scientific funding agency
CNPq through the Universal Project (grant #408919/2016-7)
and the support of NVIDIA Corporation with the donation of
the GPUs used for this research.
REFERENCES
[1] Zucchi, R. A., “Fruit flies in Brazil: Anastrepha species and their host
plants and parasitoids,” http://www.lea.esalq.usp.br/anastrepha/, 2008.
[2] Z. Bomfim, K. Lima, J. Silva, M. Costa, and R. Zucchi, “A mor-
phometric and molecular study of Anastrepha pickeli Lima (Diptera:
Tephritidae),” Neotropical Entomology, vol. 40, pp. 587–594, 2011.
[3] ——, “Morphometric and Molecular Characterization of Anastrepha
species in the spatulata Group (Diptera, Tephritidae),” Annals of the
Entomological Society of America, vol. 5, pp. 893–901, 2014.
[4] F. Faria, P. Perre, R. Zucchi, L. Jorge, T. Lewinsohn, A. Rocha, and
R. da S. Torres, “Automatic identification of fruit flies (diptera: Tephri-
tidae),” Journal of Visual Communication and Image Representation,
vol. 25, no. 7, pp. 1516–1527, 2014.
[5] P. Perre, F. A. Faria, L. R. Jorge, A. Rocha, R. S. Torres, T. Lewinsohn,
and R. A. Zucchi, “Toward an automated identification of anastrepha
fruit flies in the fraterculus group (diptera, tephritidae),” Neotropical
Entomology, vol. 0, pp. 1–5, 2016.
[6] M. M. Leonardo, S. Avila, R. A. Zucchi, and F. A. Faria, “Mid-level
image representation for fruit fly identification (diptera: Tephritidae),”
in 2017 IEEE 13th International Conference on e-Science (e-Science),
Oct 2017, pp. 202–209.
[7] M. Martineau, D. Conte, R. Raveaux, I. Arnault, D. Munier, and
G. Venturini, “A survey on image-based insect classification,” Pattern
Recognition, vol. 65, no. C, pp. 273–284, 2017.

[8] A. T. Watson, M. A. O’Neill, and I. J. Kitching, “A qualitative study
investigating automated identification of living macrolepidoptera using
the digital automated identification system (daisy),” Systematics &
Biodiversity, vol. 1, pp. 287–300, 2003.
[9] K. Russell, M. Do, J. Huv, and N. Platnick, “Introducing spida-web:
wavelets, neural networks and internet accessibility in an image-based
automated identification system,” Systematics Association, vol. 74, pp.
131–152, 2007.
[10] B. E. Boser, I. M. Guyon, and V. N. Vapnik, “A training algorithm for
optimal margin classifiers,” in Proceedings of the fifth Annual Workshop
on Computational Learning Theory, 1992, pp. 144–152.
[11] T. Arbuckle, S. Schrder, V. Steinhage, and D. Wittmann, “Biodiverssity
informatics in action: identification and monitoring of bee species using
abis.” International Symposium for Environmental Protection, pp. 425–
430, 2001.
[12] H. P. Yang, C. S. Ma, H. Wen, Q. B. Zhan, and X. L. Wang, “A tool
for developing an automatic insect identification system based on wing
outlines,” in Scientific Reports, vol. 5, 2015, p. vol. 1.
[13] S. Chen, P. Lestrel, W. Kerr, and J. McColl, “Describing shape changes
in the human mandible using elliptical fourier functions,” European
Journal of Orthodontics, vol. 22, no. 3, p. 205, 2000.
[14] C. Xie, J. Zhang, R. Li, J. Li, P. Hong, J. Xia, and P. Chen, “Automatic
classification for field crop insects via multiple-task sparse representation
and multiple-kernel learning,” Computers and Electronics in Agriculture,
vol. 119, no. C, pp. 123–132, 2015.
[15] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,”
International Journal of Computer Vision, vol. 60, no. 2, pp. 91–110,
2004.
[16] N. Dalal and B. Triggs, “Histograms of oriented gradients for human
detection,” in IEEE Conference on Computer Vision and Pattern Recog-
nition, 2005, pp. 886–893.
[17] M. Gonen and E. Alpaydin, “Multiple kernel learning algorithms,”
Journal of Machine Learning Research, vol. 12, pp. 2211–2268, 2011.
[18] F. A. Faria, J. A. dos Santos, A. Rocha, and R. da S. Torres, “A
framework for selection and fusion of pattern classifiers in multimedia
recognition,” Pattern Recognition Letters, vol. 39, pp. 52–64, 2014.
[19] S. Avila, N. Thome, M. Cord, E. Valle, and A. de A. Ara ´
ujo, “BOSSA:
Extended BoW formalism for image classification,” in IEEE Interna-
tional Conference on Image Processing, 2011, pp. 2909–2912.
[20] J. Yang, K. Yu, Y. Gong, and T. Huang, “Linear spatial pyramid match-
ing using sparse coding for image classification,” in IEEE Conference
on Computer Vision and Pattern Recognition, 2009, pp. 1794–1801.
[21] A. Lu, X. Hou, C. Lin, and C.-L. Liu, “Insect species recognition using
sparse representation,” in Proceedings of the British Machine Vision
Conference, 2010, pp. 1–10.
[22] B. Zhang, L. Xie, Y. Yuan, H. Ming, D. Huang, and M. Song,
“Deep neural network derived bottleneck features for accurate audio
classification,” in 2016 IEEE International Conference on Multimedia
Expo Workshops (ICMEW), July 2016, pp. 1–6.
[23] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[24] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking
the inception architecture for computer vision,” in Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, 2016,
pp. 2818–2826.
[25] F. Chollet, “Xception: Deep learning with depthwise separable convo-
lutions,” arXiv preprint arXiv:1610.02357, 2016.
[26] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proceedings of the IEEE conference on computer vision
and pattern recognition, 2016, pp. 770–778.
[27] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson, “How transferable are
features in deep neural networks?” in Advances in neural information
processing systems, 2014, pp. 3320–3328.
[28] Plantwise, “Empowering farmers, powering research — delivering im-
proved food security,” http://www.plantwise.org/, 2013.
[29] M. Swain and D. Ballard, “Color indexing,” International Journal of
Computer Vision, vol. 7, no. 1, pp. 11–32, 1991.
[30] E. Rosten and T. Drummond, “Machine learning for high-speed corner
detection,” in European Conference on Computer Vision, 2006, pp. 430–
443.
[31] M. D. Zeiler and R. Fergus, Visualizing and Understanding Convolu-
tional Networks. Cham: Springer International Publishing, 2014, pp.
818–833.
[32] E. Mohedano, K. McGuinness, N. E. O’Connor, A. Salvador, F. Mar-
ques, and X. Giro-i Nieto, “Bags of local convolutional features for
scalable instance search,” in Proceedings of the 2016 ACM on Interna-
tional Conference on Multimedia Retrieval, ser. ICMR ’16, New York,
NY, USA, 2016, pp. 327–331.