Insect Detection and Classiﬁcation Based
on an Improved Convolutional Neural Network
Denan Xia 1, Peng Chen 1, 2, *, Bing Wang 3, Jun Zhang 4,* and Chengjun Xie 5
1School of Computer Science and Technology, Anhui University, Hefei 230601, Anhui, China;
2Institute of Physical Science and Information Technology, Anhui University, Hefei 230601, Anhui, China
3School of Electrical and Information Engineering, Anhui University of Technology,
Ma’anshan 243032, Anhui, China; email@example.com
4School of Electrical Engineering and Automation, Anhui University, Hefei 230601, Anhui, China
5Institute of Intelligent Machines, and Hefei Institutes of Physical Science, Chinese Academy of Sciences,
Hefei 230031, Anhui, China; firstname.lastname@example.org
*Correspondence: email@example.com (P.C.); firstname.lastname@example.org (J.Z.); Tel.: +86-551-6386-1469 (P.C.)
Received: 28 August 2018; Accepted: 16 November 2018; Published: 27 November 2018
Regarding the growth of crops, one of the important factors affecting crop yield
is insect disasters. Since most insect species are extremely similar, insect detection on ﬁeld
crops, such as rice, soybean and other crops, is more challenging than generic object detection.
Presently, distinguishing insects in crop ﬁelds mainly relies on manual classiﬁcation, but this is
an extremely time-consuming and expensive process. This work proposes a convolutional neural
network model to solve the problem of multi-classiﬁcation of crop insects. The model can make full
use of the advantages of the neural network to comprehensively extract multifaceted insect features.
During the regional proposal stage, the Region Proposal Network is adopted rather than a traditional
selective search technique to generate a smaller number of proposal windows, which is especially
important for improving prediction accuracy and accelerating computations. Experimental results
show that the proposed method achieves a heightened accuracy and is superior to the state-of-the-art
traditional insect classiﬁcation algorithms.
convolutional neural network; insect detection; field crops; region proposal network; VGG19
Insects are known to be a major factor in the world’s agricultural economy, therefore it is
particularly crucial to prevent and control agricultural insects [
], through the use of programs
such as dynamic surveys and insect population management by real-time monitoring systems [
However, there are many species of insects in farmlands, which requires a lot of time for manual
classiﬁcation by insect experts [
]. It is well known that different species of insects might have
similar phenotypes, and insects often take on complicated phenotypes due to different environments
and growth periods [
]. Since people without the knowledge of entomology cannot distinguish
insect categories and the growth period of insects, it is necessary to develop more rapid and effective
approaches to tackle this problem.
The development of machine learning algorithms has provided an excellent solution for
insect image recognition [
]. Computer vision [
] and machine learning methods have achieved
great successes in vehicle identiﬁcation and pedestrian detection. Li et al. [
] combined convolutional
neural networks (CNN) with an edge boxes algorithm to accurately recognize pedestrians in images.
Several issues have to be addressed in the process of the recognition and classiﬁcation of insects,
however, which are brieﬂy described as follows:
Sensors 2018,18, 4169; doi:10.3390/s18124169 www.mdpi.com/journal/sensors
Sensors 2018,18, 4169 2 of 12
1) Quickly locate the information of an insect positioned in a complex background;
2) Accurately distinguish insect species with high similarity between intra-class and inter-class;
Effectively identify the different phenotypes of the same insect species in different growth periods.
Xie et al. [
] combined a sparse-coding technique for encoding insect images with a multiple-kernel
learning (MKL) technique to construct an insect recognition system, which achieved a mAP
(mean average precision) of 85.5% on 24 common insects in crop ﬁelds [
]. Xie’s method
requires multi-image preprocessing, however, such as image denoising and segmentation [
which expends a lot of time and technical support, so predictions on images without preprocessing
might not be satisfactory. Lim et al. adopted Alexnet and Softmax to build an insect classiﬁcation
system, which was optimized by adjusting the network architecture [
]. Yalcin et al. [
an image-based insect classiﬁcation method by using four feature extraction methods: Hu moments
(Hu), Elliptic Fourier Descriptors (EFD), Radial Distance Functions (RDF) and Local Binary Patterns
(LBP), but these images need preprocessing manually, which is undoubtedly very time consuming.
Pjd et al. [
] proposed a prototype automated identiﬁcation system which distinguishes ﬁve parasitic
wasps by identifying wing structure differences. Mayo and Watson [
] developed an automatic
identiﬁcation system using support vector machines to recognize the images of 774 live moths,
without manually specifying the region of interest (ROI). Ding and Taylor [
] proposed a neural
network model based on deep learning [
] to classify and count the number of moths and achieved
successful results. Moreover, Ding’s work showed that the model can achieve better results under ideal
experimental conditions. The quality of images, however, is often affected by sunlight, obstructions,
etc. Meanwhile, these vague images might affect the recognition and classiﬁcation of insects.
Traditional machine learning algorithms have certain limitations in the ﬁeld of image recognition.
Recently, many researchers have found that deep learning takes enormous advantage of feature
extraction in images, implemented through the adaptive learning of artiﬁcial neurons, which does
not affect the process of artiﬁcially seeking and extracting suitable features. The authors propose
a convolutional neural network model for insect recognition and classiﬁcation. The work ﬂow of this
model is roughly divided into two stages:
During the ﬁrst stage, VGG19 [
] is adopted, which is a deep network consisting of 19 layers to
extract high-dimensional features from insect images, as well as RPN, which combines highly
abstracted information trained to learn the actual locations of insects in images;
During the second stage, the feature maps are reshaped to a uniform size and converted into
a one-dimensional vector for insect classiﬁcation.
2. Materials and Methods
2.1. Dataset: Data Preprocessing and Augmentation
The dataset used in Xie’s work [
] was adopted in this work, which contains 24 common images of
insects in crop ﬁelds, such as Aelia sibirica,Atracto morphasinensis,Chilo suppressalis, etc. Figure 1shows
the scientiﬁc names and sample images of 24 insect species. To improve the generalization ability of
this model, more images collected from the Internet were used with a data augmentation technique.
Sensors 2018,18, 4169 3 of 12
Figure 1. Sample images of 24 insect species collected from crop ﬁelds.
Due to the small size of Xie’s data set, collecting new images was required. The authors manually
collected images by search engines, such as Baidu and Google, where similar images were extracted
manually. Table 1lists the information of insect species including Xie’s data set and those collected
from crop ﬁelds and the Internet. Following exclusion of some images with errors and low quality,
660 images were used in this work, where 60 images were randomly selected for the test data set
and the remaining 540 images for the training one.
Moreover, to avoid over-ﬁtting of this model, data augmentation was performed on the training
data set to increase the number of training samples. Bilinear interpolation [
] was adopted to ﬁx
images to the pixel size of 450
750, and all images were then rotated at 90
Salt and Pepper Noise [
] was also added to the images to ensure the validity of data, which randomly
changes pixel values in the images, whitening some pixel points and blackening some other pixel points.
As a result, these techniques expanded the number of training samples to eight times the original
ones. Meanwhile, an annotation ﬁle containing bounding boxes and the categories of each insect were
generated for each image.
Following the data augmentation, the training data set was expanded to 4800 images, where each
species of insect included 200 images. The number for the test data set was 480 where each
species of insect included 20 images. Thus, the insect data set “MPest” were readied for the next
Table 1. Information of 24 insect species collected from Xie’s data set, crop ﬁelds and the Internet.
Species Quantity Species Quantity Species Quantity
Aeliasibirica 66 Colposcelissignata 73 Mythimnaseparta 49
60 Dolerustritici 91
Chilosuppressalis 53 Erthesinafullo 49 Pentfaleus major 83
128 Pierisrapae 61
Cifunalocuples 47 Eurydemagebleri 42 Sitobionavenae 60
Cletus punctiger 60 Eysacorisguttiger 60 Sogatellafurcifera 71
56 Marucatestulalis 56 Tettigellaviridis 55
Sensors 2018,18, 4169 4 of 12
2.2. Deep Learning
Deep learning was proposed by Hinton [
] et al. in 2006, which is a learning model with multiple
layers of hidden perceptrons. It combines low-level features with more abstract high-level features
to discover estimable relationships in mass data sets [
]. The multi-layer perceptrons are adopted
to explore sophisticated structures with multiple abstract levels. The deep convolutional neural
network seeks hidden relationships in complex data by using the back-propagation algorithm to adjust
the parameters of neurons at each layer. Lecun proposed a multilayer neural network trained with
the back-propagation algorithm, which performed at a lower error rate, on data sets of handwritten
]. Deep learning, being better than the state-of-the-art traditional machine learning
algorithms, has greatly improved the capabilities of image recognition and more.
2.3. Overall CNN Architecture
An improved network architecture was implemented based on VGG19 [
]. Figure 2shows
the schematic diagram of this network. The 19-layer CNN network can be thought of as a self-learning
progression of local image features from low to mid to high level. The ﬁrst 16 convolutional layers of
VGG19 were adopted, which were used to extract features. Higher convolutional layers can reduce
the resolution of the feature map and extract more abstract high-level features. The Region Proposal
Network (RPN) [
] was adopted in the ﬁrst 16 layers, which can recommend the location of
insects from a feature map and remove the inﬂuence of unrelated background on classiﬁcation results.
Moreover, the last FC6 and FC7 full connection layers were used to capture complex comprehensive
feature information. This architecture is appropriate for learning local features from a complex natural
image dataset .
Figure 2. The schematic structure of the proposed detection model based on VGG19.
2.4. Region Proposal Network
Region Proposal Network [
] takes an image of arbitrary size as input, and outputs a set
of rectangle proposal boxes, where each box has an object score. To generate regional proposals,
a small network, which takes an n
nspatial window of a convolutional feature map as input,
was moved on the convolutional map of the last shared convolutional output. A 3
was selected for convolution mapping and generated a 512-dimensional feature vector. Figure 3shows
the architecture of the Region Proposal Network used in this work. At the location of each sliding
Sensors 2018,18, 4169 5 of 12
window, multiple regional proposals were generated on the current sliding window and corresponded
to various scales and aspect ratios. Here three scales and three aspect ratios were used, which resulted
in k= 9 regional proposals for each slide location. The proposals are also called anchors. The regional
proposals were then input into two full-connected layers—the bounding box regression layer (reg)
and the bounding box classiﬁcation layer (cls).
Figure 3. Region Proposal Network (RPN).
To train the Region Proposal Network, for one image, the target proposal area in the image
was assigned a plurality of binary class labels (insects or backgrounds), while the remaining
area was discarded. The proposal region was assigned to a positive label, if it had the highest
intersection-over-union (IoU) overlap ratio with the ground truth box; otherwise, it was assigned to
a negative label if the IoU of the proposal region was lower than the IoU threshold of all ground-truth
boxes. The IoU ratio is deﬁned as follows:
IoU =area(Bpest ∩Bground)
area(Bpest ∪Bground), (1)
)represents the intersection area of the insect proposal box and ground truth
box, and area(Bpest ∪Bground)denotes their union area.
Given one image with insects, it is hoped to extract a fixed-length feature vector for an insect from
a complex image background by convolution operation. The region of interest (ROI) pooling was adopted
to convert insect-like regions into a fixed spatial size which facilitates the generation of the same
dimensional feature vectors, because these insect-like regions have different sizes. Then, each ROI feature
map was input into the FC6 and FC7 fully connected layers whose output was a 4096-dimension feature
vector including the location and category information of the target. The feature vector was then input
into the Softmax layer to identify insects and estimate the insect-like regions simultaneously.
Training Region Proposal Network
Each input image was resized to 450
750 pixels as previously discussed. The feature
map was obtained from the input image by the convolution operation of this network.
However, convolution operation lead to huge mathematical operations, which might result in
the problem of gradient explosion or gradient disappearance. Therefore, it was necessary to add
a rectification non-linearities (Relu) layer to activate or suppress the output characteristic diagram
of each convolutional layer. A max pooling layer was added after the second, fourth, seventh,
Sensors 2018,18, 4169 6 of 12
eleventh and fifteenth convolutional layers. Moreover, to prevent network over-fitting, the pre-trained
VGG19 model was adopted to initialize the first sixteen convolutional layers in this model.
2.5. Loss Function
Regarding loss function of this model, different loss functions were employed for the bounding
box regression layer and the bounding box classification layer. The former layer can make an insect-like
category score Pfor each predicted region, while the latter one can output a coordinate vector
loc = (x,y,m,n)
for each predicted region, where xand ydenote the horizontal and vertical coordinates
of the predicted region, respectively, while mand ndenote the width and height of the predicted region.
Subsequent to defining the loss functions for classification and regression, they were combined by
following Girshick’s multi-task loss rule  thus, the whole loss function of this proposed model is:
βNreg , (2)
is the predicted probability of anchor ibeing an object; the ground-truth label
is 1 if
the region box is positive, otherwise
denotes the four parameterized coordinates of
the predicted bounding box; and
is the ground-truth box associated with a positive region box.
is the balancing parameter. N
is the mini-batch size and N
is the number of anchor locations.
is the classiﬁcation loss function over two classes (object or background), which is written as follows:
i) = Log[αiα∗
To determine the regression loss, Lreg denotes a smooth L1loss as:
i) = f(si,s∗
where f(x) = (0.5x2if |x|<1
2.6. Training Overall Model
The proposed model was implemented in combination with VGG19 and RPN models. The weights
of the network were initialized by the pre-training VGG19 model, while the initial learning rate
was 0.001, the momentum was 0.9, and the weight decay was 0.0005. The weights were updated by
stochastic gradient descent trick. VGG19 and RPN can be alternately trained to optimize the model
rather than training two separate networks. During the ﬁrst step, the RPN network was trained by
learning an image with the location bounding box of insect-like. Next, the predicted region generated
by RPN was input to VGG19 to train itself. Finally, both RPN and VGG19 were trained jointly
and ﬁne-tuned by ﬁxing shared convolutional network layers. Figure 4illustrates the ﬂowchart of
insect recognition and classiﬁcation.
Sensors 2018,18, 4169 7 of 12
Flowchart of insect recognition and classiﬁcation. Abundant images were obtained by taking
photos in crop ﬁelds and collecting images by Baidu and Google online, in which insect images are
original and the quality of insect images are uneven. Then, image preprocessing and data augmentation
were applied to form our own dataset “MPest”. The model, which was trained on the dataset “MPest”,
can effectively help to recognize insects and diseases.
3. Experiments and Results
This section shows the evaluation of this method for insect recognition and the analysis of
this proposed model with different parameters in detail. All experiments were implemented on
the framework of Caffe [
] and mean Average Precision (mAP) in Everingham’s work [
] was used
as an evaluation metric.
3.1. Effects of Feature Extraction Network
To evaluate the effect of feature extraction on the performance of this model, several feature
extraction networks were implemented on the dataset “MPest”, which contains VGG16 with different
convolutional features. Generally, the greater number of convolutional layers the model has, the more
complex features the model can learn from the images. All of the comparison methods performed
160,000 iterations, with the initialized learning rate of 0.001. Figure 5shows the performance
comparison of the three methods on this dataset. It can be seen that the VGG16 network achieved
good performance, but the proposed network achieved an improvement of 3.72% over VGG16.
Regarding the ZF network, it consists of a simple architecture of only ﬁve convolutional layers
7 ﬁlters, which results in quick feature map shrink and, thus, cannot extract effectively
the multi-faceted features of insects from images. Both VGG16 and VGG19 belong to deep neural
networks, where the former contains sixteen convolutional layers and the latter contains nineteen layers.
The model with high-level convolution layers can effectively enhance the ability of the proposed model
to extract information on insect characteristics under complex backgrounds. Moreover, a 3
was implemented in VGG because multiple 3
3 small ﬁlters have more nonlinearity than a 7
ﬁlter, which makes the decision function more decisive.
Sensors 2018,18, 4169 8 of 12
Figure 5. Comparison of different feature extraction methods.
Figure 6shows the convolution map of three feature extraction networks on our dataset
“MPest”. The feature maps of the ZF network have a high loss rate of pixel information and the ZF
network blurred images so that it was hard to distinguish insects in images. The feature maps
of VGG16 showed that the pixel information of images was retained perfectly by using the small
ﬁlters. Although the feature maps of VGG19 and VGG16 are almost the same, the more abstract
high-dimensional information was retained since more convolutional layers were applied in VGG19.
Visualization of feature maps of different feature extraction networks. (
): ZF Net; (
3.2. Effects of Iou Threshold
The goal of this proposed training task is to reduce the difference between predicted region
and ground truth, therefore reducing the inﬂuence of unrelated background noise in the predicted
region is immensely signiﬁcant. The IoU threshold score was varied from 0.3 to 0.8 in step sizes of 0.1.
Figure 7shows the mAP curve increased gradually with the IoU threshold, from 0.3 to 0.5.
Sensors 2018,18, 4169 9 of 12
Figure 7. Effects of IoU threshold (VGG19).
That is to say, the increase of the threshold results in abandoning more predicted regions
which overlap less with the ground truth. The mAP curve reaches maximum value when the threshold
is set to 0.5, where 89.22% of the ground truth are detected successfully. Starting from 0.6 to 0.8,
the curve declines slowly. The larger the value of IoU is, the more predicted regions the model
abandoned in the regional proposal stage. Therefore, the lower threshold results in an excessively
small overlap area between the predicted region and ground truth, where more backgrounds were
present in the classiﬁcation task. Higher thresholds lead to larger discarded prediction regions,
which results in an unsuccessful training model.
3.3. Effects of Learning Rate
Learning rate is an important hyperparameter that controls the update speed of network weights.
The effect of the hyperparametric learning rate was investigated from 0.0006 to 0.0014. The learning
rate technique of this work is not adaptive. This proposed model ﬁrst sets an initial learning rate
and then decreases it 10 times every 20,000 iterations. Figure 8shows the mAP of this model with
respect to the learning rate.
Figure 8. Effects of Learning Rate (VGG19).
Sensors 2018,18, 4169 10 of 12
Looking at Figure 8, it can be seen that, as the learning rate increases, the error of the model
gradually decreases. Moreover, a small learning rate leads to a slow update speed of weights of
the model and, subsequently, the convergence of the model is not ideal. Therefore, as the learning
rate increases, the experimental error of the model gradually decreases. When the learning rate
exceeds 0.001, the mAP of the proposed model begins to decrease. A higher learning rate means
that the convergence speed of the model is too fast, which results in a larger loss value than expected
in the iterative process and makes the model over-ﬁtting.
3.4. Performance Comparison with Other Methods
Table 2shows the performance comparison with other state-of-the-art methods to this study’s test
set, such as Single Shot Multibox Detector (SSD) and Fast Region-based Convolutional Neural Network
(RCNN). It is known that SSD is a prominent algorithm in which the recurrent feature-pyramid structure
is employed to detect images. Several separate predictors were adopted to perform classification
and regression tasks at the same time in the multi-feature mapping of the network and the processing of
the target detection problem using multi-feature information. The SSD yielded the best performance
when it ran 30,000 iterations with a learning rate of 0.001. Obviously, the proposed method outperformed
the SSD model and achieved an improvement in mAP of 3.73%. The inference time of the proposed
method was the least among the three methods, about 0.083 s. Moreover, the training time of SSD took
38 hours, which was longer than the proposed method.
Table 2. Comparison with Other Methods.
Method mAP Inference Time(s)/Per Image Training Time(h)
0.8922 0.083 11.2
SSD 0.8534 0.120 38.4
Fast RCNN 0.7964 0.195 70.1
Fast RCNN is a regional proposal and target classiﬁcation algorithm that adopts selective
search techniques to generate proposal windows. Table 2shows Fast RCNN achieved a mAP
of 0.7964 after 60,000 iterations. However, Fast RCNN took about 70 hours for the training model while
the detection time of each image was about 0.195 s. Therefore, fast RCNN requires more computational
resources and time for insect detection than this proposed method.
Attempts were made to compare the proposed model with the latest network modules and models.
Inception was added to this model, for instance, or different convolution layers of this model
were replaced, but experimental results showed that it changed the accuracy of the model little.
Moreover, Resnet (Residual Neural Network) could not gain a satisfactory result because the pixel size
of images in this work was too large.
Moreover, the differences between this proposed model and other methods can be summarized in
two aspects. First, for the insect data set, insect images were collected under ﬁeld conditions rather
than under ideal conditions, which offers the proposed model stronger anti-interference capability.
Second, for insect recognition, this proposed method actually can locate insects in images, while most
other methods only implemented image classiﬁcation. To conclude, this proposed method effectively
can moderate the distraction of human factors and artiﬁcial burdens in processing data set.
4. Conclusions and Future Work
A target recognition method based on improved VGG19 for quickly and accurately detecting
insects in images was proposed. Since Caffe library provides a pre-trained VGG19 model,
whose architecture has achieved a successful balance in feature extraction, the current authors
ﬁne-tuned the pre-trained model to train this study’s ideal model. The experimental results on
the current dataset “MPest” showed that this method is faster and more accurate than existing methods.
Sensors 2018,18, 4169 11 of 12
However, there are still some issues in this proposed method, such as target detection error.
Therefore, this method performance can be further improved as follows:
1) The insect database needs to be augmented, which can be manually collected in the future;
2) More appropriate models to extract helpful insect-like areas from images should be tried;
Regarding the classiﬁcation task, the classiﬁcation of insects needs to be more detailed,
and the periods of insect growth should be divided. Workers will implement different pest
control measures according to the period of insect growth.
The source codes are available online at http://deeplearner.ahu.edu.cn/web/cnnPest.htm.
Conceptualization, P.C.; Data curation, D.X. and C.X.; Funding acquisition, P.C.;
Investigation, B.W.; Methodology, J.Z.; Project administration, P.C.; Software, J.Z. and C.X.; Supervision, P.C.;
Validation, B.W.; Writing–original draft, D.X.; Writing–review & editing, P.C., B.W. and C.X..
This research was funded by the National Natural Science Foundation of China (Nos. 61672035,
61300058, 61872004, 61472282 and 31401293).
Conﬂicts of Interest: The authors declare no conﬂict of interest.
Estruch, J.J.; Carozzi, N.B.; Desai, N.; Duck, N.B.; Warren, G.W.; Koziel, M.G. Transgenic plants: An emerging
approach to pest control. Nat. Biotechnol. 1997,15, 137–141. [CrossRef] [PubMed]
Fedor, P.; Vaˇnhara, J.; Havel, J.; Malenovsky, I.; Spellerberg, I. Artiﬁcial intelligence in pest insect monitoring.
Syst. Entomol. 2009,34, 398–400. [CrossRef]
Hanysyam, M.N.M.; Fauziah, I.; Khairiyah, M.H.S.; Fairuz, K.; Rasdi, Z.M.; Zfarina, M.Z.N.; Elﬁra, S.E.;
Ismail, R.; Norazliza, R. Entomofaunal Diversity of insects in FELDA Gunung Besout 6, Sungkai, Perak.
In Proceedings of the 2013 IEEE Business Engineering and Industrial Applications Colloquium (BEIAC),
Langkawi, Malaysia, 7–9 April 2013; pp. 234–239.
4. Gaston, K.J. The Magnitude of Global Insect Species Richness. Conserv. Biol. 2010,5, 283–296. [CrossRef]
Siemann, E.; Tilman, D.; Haarstad, J. Insect species diversity, abundance and body size relationships.
Nature 1996,380, 704–706. [CrossRef]
Zhang, H.; Huo, Q.; Ding, W. The application of AdaBoost-neural network in storedproduct insect
classiﬁcation. In Proceedings of the IEEE International Symposium on It in Medicine and Education,
Xiamen, China, 12–14 December 2009; pp. 973–976.
Xie, C.; Zhang, J.; Li, R.; Li, J.; Hong, P.; Xia, J.; Chen, P. Automatic classiﬁcation for ﬁeld crop insects
via multiple-task sparse representation and multiple-kernel learning. Comput. Electron. Agric.
119, 123–132. [CrossRef]
Xie, C.; Wang, R.; Zhang, J.; Chen, P.; Li, R.; Chen, T.; Chen, H. Multi-level learning features for automatic
classiﬁcation of ﬁeld crop pests. Comput. Electron. Agric. 2018,152, 233–241. [CrossRef]
Szeliski, R. Computer Vision: Algorithms and Applications; Springer: London, UK, 2010; Volume 21,
Li, H.; Wu, Z.; Zhang, J. Pedestrian detection based on deep learning model. In Proceedings of
the International Congress on Image and Signal Processing, Biomedical Engineering and Informatics, Datong,
China, 15–17 October 2017; pp. 796–800.
Everingham, M.; Eslami, S.M.A.; Gool, L.V.; Williams, C.K.I.; Winn, J.; Zisserman, A. The Pascal, Visual Object
Classes Challenge: A Retrospective. Int. J. Comput. Vis. 2015,111, 98–136. [CrossRef]
Lin, D.; Lu, C.; Huang, H.; Jia, J. RSCM: Region Selection and Concurrency Model for Multi-class Weather
Classiﬁcation. IEEE Trans. Image Process. 2017,26, 4154–4167. [CrossRef] [PubMed]
Keys, R. Cubic convolution interpolation for digital image processing. IEEE Trans. Acoust. Speech Signal Process.
2003,29, 1153–1160. [CrossRef]
Lim, S.; Kim, S.; Kim, D. Performance effect analysis for insect classiﬁcation using convolutional
neural network. In Proceedings of the 2017 7th IEEE International Conference on Control System,
Computing and Engineering (ICCSCE), Penang, Malaysia, 24–26 November 2017; pp. 210–215.
Sensors 2018,18, 4169 12 of 12
Yalcin, H. Vision based automatic inspection of insects in pheromone traps. In Proceedings of the International
Conference on Agro-Geoinformatics, Istanbul, Turkey, 20–24 July 2015; pp. 333–338.
Pjd, W.; O’Neill, M.A.; Gaston, K.J.; Gauld, I.D. Automating insect identiﬁcation: Exploring the limitations of
a prototype system. J. Appl. Entomol. 2010,123, 1–8.
Mayo, M.; Watson, A.T. Automatic Species Identiﬁcation of Live Moths. Knowl.-Based Syst.
Ding, W.; Taylor, G. Automatic moth detection from trap images for pest management. Comput. Electron. Agric.
2016,123, 17–28. [CrossRef]
19. Lecun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015,521, 436. [CrossRef] [PubMed]
Simonyan, K.; Zisserman, A. Very Deep Convolutional Networks for Large-Scale Image Recognition.
Kirkland, E.J. Bilinear Interpolation. Advanced Computing in Electron Microscopy; Springer: Boston, MA, USA,
2010; pp. 261–263.
Chan, R.H.; Ho, C.W.; Nikolova, M. Salt-and-pepper noise removal by median-type noise detectors
and detail-preserving regularization. IEEE Trans. Image Process. A Publ. IEEE Signal Process. Soc.
Du, X.; Cai, Y.; Wang, S.; Zhang, L. Overview of deep learning. In Proceedings of the 2016 31st
Youth Academic Annual Conference of Chinese Association of Automation (YAC), Wuhan, China,
11–13 November 2016; pp. 159–164.
Lecun, Y.; Bottou, L.; Bengio, Y.; Haffner, P. Gradient-based learning applied to document recognition.
Proc. IEEE 1998,86, 2278–2324. [CrossRef]
Ren, S.; He, K.; Girshick, R.; Sun, J. Faster R-CNN: Towards real-time object detection with region
proposal networks. In International Conference on Neural Information Processing Systems; MIT Press:
Cambridge, MA, USA, 2015; pp. 91–99.
Sommer, L.; Schumann, A.; Schuchert, T.; Beyerer, J. Multi Feature Deconvolutional Faster R-CNN for Precise
Vehicle Detection in Aerial Imagery. In Proceedings of the IEEE Winter Conference on Applications of
Computer Vision. IEEE Computer Society, Lake Tahoe, NV, USA, 12–15 March 2018; pp. 635–642.
Razavian, A.S.; Azizpour, H.; Sullivan, J.; Carlsson, S. CNN features off-the-shelf: An astounding baseline for
recognition. In Proceedings of the 2014 IEEE Computer Society Conference on Computer Vision and Pattern
Recognition (CVPR 2014), Columbus, OH, USA, 24–27 June 2014.
Girshick, R. Fast R-CNN. In Proceedings of the 2015 IEEE International Conference on Computer Vision
(ICCV), Santiago, Chile, 7–13 December 2015; pp. 1440–1448.
Jia, Y.; Shelhamer, E.; Donahue, J.; Karayev, S.; Long, J.; Gir-shick, R.; Guadarrama, S.; Darrell, T. Caffe:
Convolutional architecture for fast feature embedding. In Proceedings of the ACM International Conference
on Multimedia, Orlando, FL, USA, 3–7 November 2014; pp. 675–678.
2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).