Neurocomputing 267 (2017) 378–384

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

Identification of rice diseases using deep convolutional neural

networksR
Yang Lu a,b,∗, Shujuan Yi a, Nianyin Zeng c, Yurong Liu d,e, Yong Zhang f
a
College of Information Technology, Heilongjiang Bayi Agricultural University, Daqing, Heilongjiang 163319, PR China
b
Fujian Provincial Key Laboratory of Information Processing and Intelligent Control, Minjiang University, Fuzhou, PR China
c
Department of Instrumental and Electrical Engineering, Xiamen University, Xiamen, Fujian 361005, PR China
d
Department of Mathematics, Yangzhou University, Yangzhou, Jiangsu 225002, PR China
e
Communications Systems and Networks (CSN) Research Group, Faculty of Engineering, King Abdulaziz University, Jeddah 21589, Saudi Arabia
f
College of Electrical Science, Northeast Petroleum University, Daqing, Heilongjiang 163318, PR China

a r t i c l e i n f o a b s t r a c t

Article history: The automatic identification and diagnosis of rice diseases are highly desired in the field of agricultural
Received 6 February 2017 information. Deep learning is a hot research topic in pattern recognition and machine learning at present,
Revised 27 May 2017
it can effectively solve these problems in vegetable pathology. In this study, we propose a novel rice dis-
Accepted 17 June 2017
eases identification method based on deep convolutional neural networks (CNNs) techniques. Using a
Available online 1 July 2017
dataset of 500 natural images of diseased and healthy rice leaves and stems captured from rice exper-
Communicated by Zidong Wang imental field, CNNs are trained to identify 10 common rice diseases. Under the 10-fold cross-validation
strategy, the proposed CNNs-based model achieves an accuracy of 95.48%. This accuracy is much higher
Keywords:
Identification of rice diseases than conventional machine learning model. The simulation results for the identification of rice diseases
Convolutional neural networks show the feasibility and effectiveness of the proposed method.
Deep learning © 2017 Elsevier B.V. All rights reserved.
Image recognition

1. Introduction diseases remotely. The disadvantage is that not only the judgement
is easy to be wrong, but also the efficiency is low.
Rice is one of the most important food crops in China and even In order to improve the accuracy and rapidity of the diagnosis
in the world. Rice diseases have a devastating effect on rice pro- results, many researchers have studied the automated rice diseases
duction. Also, it is a major threat to food security. Thus, the diag- diagnosis based on pattern recognition and machine learning. Such
nosis and identification of rice diseases play a very important role as, using pattern recognition techniques [25], support vector ma-
in ensuring the high yield, high quality and high efficiency of rice. chine [16], digital image processing techniques [2] and computer
The traditional method of detecting rice diseases requires lots of vision[1]. Meanwhile, these advanced techniques are not only ap-
experts’ experience and knowledge. With the development of com- plied to the diagnosis of rice diseases, but also to other crops, such
puter and internet technology, farmers can search the rice diseases as wheat [20], maize [34], cotton [29], tomato [5], etc.
images database or consult the plant pathologists to judge rice The past 10 years have witnessed successful applications of
deep convolutional neural networks (CNNs) in diverse fields in-
cluding image classification [6,21], video classification [18], traffic
R
This work was supported in part by the National Natural Science Foundation of sign recognition [17] and human action recognition [15], etc. The
China under grants 61374127 and 61422301, the Outstanding Youth Science Foun- study of CNNs and related research have therefore gained persis-
dation of Heilongjiang Province under grant JC2015016, the Natural Science Foun- tent research interest since the early 1990s, see [10,12] and the
dation of Heilongjiang Province under grant F201428, the Science and Technology
references therein. During this period, Hinton and co-workers have
Research of Agricultural Bureau in Heilongjiang Province under grant HNK125B-04-
03, China Postdoctoral Science Foundation under grant 2016M591560, Heilongjiang
done widely meaningful and fundamental research on deep neu-
Postdoctoral Financial Assistance under grant LBH-Z15185, Heilongjiang Bayi Agri- ral network to improve algorithm performance and optimize archi-
cultural University Foundation under grant XA2016-05 and Fujian Provincial Key tecture, see e.g. [3,23,24,27], and the references cited. A number
Laboratory of Information Processing and Intelligent Control (Minjiang University) of research have been made for CNNs to improve the original ar-
Opening Fund under grant MJUKF201729.
∗
chitecture of Krizhevsky et al. [21]. For example, Zeiler and Fergus
Corresponding author at: College of Information Technology, Heilongjiang Bayi
Agricultural University, Daqing, Heilongjiang 163319, PR China.
[33] used stochastic pooling for regulation of CNNs and Simonyan
E-mail address: [email protected] (Y. Lu). and Zisserman [31] proposed very deep CNNs to 16–19 layers,

http://dx.doi.org/10.1016/j.neucom.2017.06.023
0925-2312/© 2017 Elsevier B.V. All rights reserved.
 Y. Lu et al. / Neurocomputing 267 (2017) 378–384 379

which achieved the state-of-art accuracy on ImageNet Large-Scale

Visual Recognition Challenge (ILSVRC) [26]. Especially, in March
2016, AlphaGo beat Lee Sedol in a five-game match, the first time a
computer Go program has beaten a 9-dan professional, and it was
a major milestone in artificial intelligence research [30,32]. LeCun
et al. pointed out that deep convolutional neural networks have
brought about breakthroughs in processing images, video, speech
and audio in a survey paper published in Nature in 2015 [22].
Recently, several studies on automated plant disease diagno-
sis have been conducted using deep learning techniques [19,35].
Fig. 1. The structure of deep convolutional neural networks.
Kawasaki et al. proposed a system based on CNNs to recognize cu-
cumber leaf disease; it achieves an accuracy of 94.9% in distin-
guishing among melon yellow spot virus, zucchini yellow mosaic Table 1
virus and a non-diseased class. Sanyal et al. have identified rice Related parameters of CNNs.

disease of brown spot and blast diseases by using color texture of Layer name Function Weight filter sizes Output tensor
rice leaf images [28]. Input 3 × 512 × 512
As we all know that the machine learning technology is essen- Conv1 Convolutional 16 × 16 3 × 224 × 224
tial in the intelligent diagnosis of crop diseases. However, there ex- Pool1 Stochastic pooling 5×5 32 × 112 × 112
ist some defects and difficulties in the above research, involving Conv2 Convolutional 9×9 64 × 112 × 112
Pool2 Stochastic pooling 5×5 64 × 56 × 56
the recognition rate is higher for specific samples under certain
Conv3 Convolutional 5×5 128 × 28 × 28
circumstances. In the diagnosis model, some parameters are not Pool3 Stochastic pooling 5×5 128 × 14 × 14
optimal, the convergence speed of training algorithm is slow, and Output Softmax regression 10 × 1 × 1
it is easy to fall into local minima, and so on.
So far, no research has been published which explores deep
convolutional neural networks for rice diseases identification. The
goal of this research is to construct deep convolutional networks from a reduced data set applying convolution and pooling. The
model to achieve fast and accurate automated recognition by us- stochastic pooling method is applied in the proposed CNNs model.
ing rice diseases images. The 10 common rice diseases include For it is simple enough to randomly select elements in feature map
rice blast (RB), rice false smut (RFS), rice brown spot (RBS), rice according to their probability values, that is, the elements with
bakanae disease (RBD), rice sheath blight (RSHB), rice sheath rot large probability are easy to be chosen. Unlike max-pooling, the
(RSR), rice bacterial leaf blight (RBLB), rice bacterial sheath rot maximum element always chosen, and mean-pooling, the average
(RBSR), rice seeding blight (RSEB) and rice bacterial wilt (RBW). value of the elements is chosen. Also, it strengthened generaliza-
The key motivation for developing the deep convolutional net- tion ability of the proposed CNNs model. Third, softmax regression
works model for rice disease is to provide the farmers an easy-to- learning algorithm can solve multi-classification problem. Finally,
use system to detect early-stage infections by using common digi- we can distinguish between 10 common different rice diseases.
tal camera. Second, extracting effective features for identifying rice The remainder of this paper is organized as follows. In
diseases is a critical but challenging task, and CNNs are highly ex- Section 2, we describe the architecture and learning algorithm of
pected to be automated feature learning from the raw inputs in CNNs. Section 3 introduces the rice disease identification method.
a systematic way. Through the deep architecture, the learned fea- The application of the developed CNNs to the problem of rice dis-
tures are deemed as the higher level abstract representation of low eases recognition is presented in Section 4. Conclusions are given
level raw rice disease images. In addition, to improve diagnostic in Section 5.
results, the CNNs is regarded as one of the best classifications in
pattern recognition tasks. So we develop the deep convolutional
neural networks model for rice diseases identification in this pa-
per. 2. Architecture of deep convolutional neural networks model
In this paper, we present a novel rice diseases identification
method based on deep convolutional neural networks. The pro- Inspired by classical and successful LeNet-5, AlexNet CNNs ar-
posed CNNs-based model can get higher classification ratio. On chitecture and its performance improved by Ciresan et al. [6],
the basis of an analysis of structure and parameters in CNNs, the Krizhevsky et al. [21] and LeCun et al. [23], we describe a kind
gradient-descent algorithm can be applied to train CNNs. A total of our multi-stage-CNNs configuration. The CNNs-based model in-
of 500 rice leaves and stem images are pre-processed first, then cludes convolution layer, stochastic pooling layer, and softmax
the processed images are used to train CNNs. This CNNs model layer. An illustration and related parameters are shown in Fig. 1
can improve the convergence speed while training the parame- and Table 1.
ters in CNNs, and obtain a higher recognition accuracy than the The size of the input image is set to be 224 × 224 × 3 according
conventional model. The main contribution of this paper is mainly to the experience, which can be divided by 2, and computing capa-
twofold. (1) Convolutional neural network is first applied to the bility of the computer. It is composed of many slices in the depth
issue of the identification of rice diseases. Note that the proposed direction. One slice corresponds to many neurons, the weight in
method can correctly and effectively recognize the 10 common rice neurons can be thought as the convolution kernel, that is a square
diseases. (2) Experiment results show that the CNNs method can filter, such as 16 × 16, 9 × 9 and 5 × 5. These neurons respectively
not only improve the convergence speed, but also obtain a higher correspond to a local area in the image, and it is used to extract
recognition accuracy than other models: (1) standard BP algorithm; the feature of the region. Assume that the size of the input im-
(2) support vector machine (SVM); (3) particle swarm optimization age is W, the size of the convolution kernel is F, the mobile stride
(PSO). of the convolution kernel is S, generally S = 2, padding P is used
The advantages of the proposed CNNs are that images can be to fill in the input image boundary, usually P = 0, so the size of
input directly to the model. First, sparse-auto encoding is used to the image after convolution is (W − F + 2P )/S + 1. Then we get the
learn the features from images. Second, we can classify images output tensor.
380 Y. Lu et al. / Neurocomputing 267 (2017) 378–384

2.1. Convolutional layer 2.3. Softmax regression

The proposed CNNs are of hierarchical architecture, which con- Softmax regression is applied in multi-class classification prob-
tain three convolutional layers. The first convolutional layer is used lem. The hypothesis function takes the form:
to extract different low-level features from the input image, such 1
as edges, lines and corners. The other two can get high-level fea- hθ ( x ) = (5)
1 + exp(−θ T x )
tures.
Each output map feature combines multiple input maps with Our goal is to train θ to minimize the cost function J(θ ).
convolutions. Generally, the output can be denoted by the follow-

1   (i )
m l
ing formula: J (θ ) = − l {y = j}log p(y(i ) = j|x(i ) ; θ ) (6)
m
  i=1 j=0

xj = f xi −1 ∗ ki j + bj (1) For training dataset { ( x ( 1 ) , y ( 1 ) ), . . . , ( x ( m ) , y ( m ) )}, yi ∈
i∈M j
{1, 2, . . . , k}. In Softmax regression, the probability of classify-
ing x as category j is:
where  represents the th layer, kij represents convolutional ker- T (i )
eθ j x
nel, bj represents bias and Mj represents a set of input maps. The p ( y ( i ) = j | x ( i ) ; θ ) = k (7)
θlT x(i)
detailed implementation of the CNNs can use a sigmoid function l=1 e
or a tanh function, also an additive bias can be applied in it. For Supervised learning algorithm is used to train the network to
example, the value of unit at position (x, y) in the jth feature map learn. The internal representation reflects the similarity among
xy
and the ith layer, denoted as vi j , is given by training samples. We design 10 classifications for rice disease iden-
  tification. In order to know the feature representation learned by
Pi −1 Q j −1
 the deep concolutional neural network, the maximized activation
xy
v = sigmoid bi j +
ij
wipqj ((xi−1
v + p)(y+q )
) (2) neuron represents the recognized rice disease. At last, we visual-
p=0 q=0 ize the image features by averaging image patches that are related
with the neurons with stochastic response in an upper-layer.
where sigmoid(·) is the sigmoid function, bij is the bias for the fea-
pq
ture map, Pi and Qj are the height and width of the kernel, and wi j 2.4. Training algorithm
is the kernel weight value at the position (p, q) connected to the (i,
j) layer. The parameters of CNNs, such as the bias bij and the ker- Also, we use back gradient-descent algorithm to train CNNs. It
pq includes two stages of feedforward pass and back propagation pass.
nel weight wi j , are usually trained using unsupervised approaches
[4,23]. In feedforward pass stage, we consider a multiclass problem with c
classes and N training samples. The squared-error function is given
by
2.2. Stochastic pooling layer c  2
1  n
N
EN = tk − ynk (8)
2
We use a stochastic-pooling layer to substitute a sub-sampling n=1 k=1
layer in CNNs [23]. Stochastic-pooling layer is used to reduce vari- where tkn is the kth dimension of the nth pattern’s corresponding
ance, it can compute the max value of a particular feature over a label, and ynk is the value of the kth output layer unit in response
region of the image. Suppose the image features have small trans- to the nth input pattern.
lations, it will ensure that we can obtain the same result. Stochas- Let  represent the current layer, L denote the output layer and
tic pooling not only combines the advantages of meaning-pooling the input layer be denoted by l. The output of the current layer is
and max-pooling but also prevents over-fitting. It plays a very im- defined to be
portant role in rice diseases classification and detection.  
In stochastic pooling, first, we need to compute the probabilities x = sigmoid W  x−l + b (9)
p for each region j according to the Eq. (3)
where W represents weight value matrix and b represents bias.
αi In back propagation pass stage, from higher layers to lower lay-
pi =  (3)
k∈R j αk ers, the back propagated error can use the following recurrence
function:
where Rj is pooling region j in feature map c, and i is the index of  T  
each element within it. δ  = W +1 δ +1 f  W  x−l + b (10)
Then we sample from the multinomial distribution based on p
to pick a location l within the region. The stochastic pooling oper- where “ ” represents elements multiplication. Finally, we will up-
ation can be denoted by Stochastic(s), for each feature map c, the date the weight according to the following formulas:
Stochastic(s) is formulated by: ∂E
  = x−1 (δ  )T (11)
∂W 
p,q = Stochastic(m,n,i, j )∈ p αm,n u (i, j )
α l,k l−1,k
(4)
∂E
W  = −η (12)
∂W 
where α l,k
p,q is the activation of the neuron at spatial coordinate
(p, q) in feature map k in lt h layer, u(i, j) is the weighing window 3. Rice disease identification method
function.
The advantages of this implementation are that stochastic pool- 3.1. Rice diseases images data acquisition and processing software
ing layer not only makes the CNNs have faster convergence, but
also improves generalization ability when processing superior in- Rice diseases images database is created, which consists of a
variant features. total of 500 rice diseases images. Some images of rice diseases
 Y. Lu et al. / Neurocomputing 267 (2017) 378–384 381

Color
images
training
sample PCA &
Whitening Training
Color
Scale Random images test Feature
Rice The dataset of
Normalization sampling samples Map
Disease Normalized
Images Rice Disease
Mean Gray PCA &
Database Images
Normalization images Whitening
Test
training Feature
sample Map

Gray Fig. 3. Rice sheath blight recognition image.

images test
samples

Fig. 2. Rice diseases images pre-processing procedure.

are captured from Heilongjiang Academy of Land Reclamation Sci-

ences, China. The dataset consists of 500 common rice disease im-
ages. The 10 common rice diseases include rice blast (RB), rice false
smut (RFS), rice brown spot (RBS), rice bakanae disease (RBD), rice
sheath blight (RSHB), rice sheath rot (RSR), rice bacterial leaf blight
(RBLB), rice bacterial sheath rot (RBSR), rice seeding blight (RSEB)
and rice bacterial wilt (RBW). Canon EOS 5D Mark III (EF 24–70
mm F2.8L II USM) digital color camera has been used in capturing
rice disease images with resolutions of 5760 × 3840 pixels. Some
images of rice diseases are collected from agricultural pest and
insect pests picture database (http://bcch.ahnw.gov.cn/Right.aspx).
Some rice diseases images are scanned from book, titled with ‘Rice
Diseases and Insect Pests Atlas and Control Technology’ [7,36].
Fig. 4. Rice disease image patches.
The images are processed and analyzed by using digital image
processing toolbox and stanford_dl_ex-masterCNNs toolbox in Mat-
lab R2012a software.

3.2. Rice diseases pre-processing

In order to reduce the running time of matlab program, the

size of the rice disease images are compressed from 5760 × 3840 to
512 × 512. CNNs is a part-connected network, not a full-connected
network. In full-connected network, if all nodes in input layer
would connect all the nodes in hidden-layer, the parameters that
be learned will become very much, and the training time will be-
come very long. In part-connected network, the nodes in hidden
layer just connect partial nodes in input layer. This method simu-
lates the visual cortex in the human cerebral cortex, and different
positions only respond to the local region. Since natural images are
stable, that is, the statistical features of a part of an image are simi- Fig. 5. The corresponding feature map of rice disease image patches.
lar to those of other parts, so the characteristics of the parts which
we learn are also applicable to other parts. Also, due to that all
Then apply ZCA-Whitening method to remove correlation be-
the rice diseases images are color images, the stationarity property
tween data X according to the following formula:
does not hold across color channels, then we rescale the data in
[0, 1] and apply PCA and Whitening to get training feature and xrot
XZCA,Whitening = U · · UT (14)
testing feature. The pre-processing procedure is shown in Fig. 2. S+ε
First, we turn the image into fixed-size 512 × 512 RGB image in
where U = svd(S ), xrot = U T · x, the reasonable range of parameter
order to reduce the dimension of training data and keep the details
ε is in 0.01–0.1.
of the input image. Let n-dimensional vector x represent any one
image, n is the product of the image width and height. The pre-
processing on the N images in rice diseases database is computed 4. Rice disease recognition simulation examples
from each pixel by the following formula:
The proposed CNNs model is applied to rice disease recognition
x˜i − μ problem. We use database of 500 images labeled of 10 kinds of
xi = √ (13) rice diseases. One of the rice disease images is shown in Fig. 3.
S+8 Randomly select 10,0 0 0 12 × 12 patches from the 500 natural
N N images, according to the pre-processing procedure introduced in
where μ = 1
N j=1 x j, S = 1
N k=1 (xk − μ )(xk + μ )T , x˜i is the esti- Section 3.2, we get one of the rice disease image patches as shown
mate of xi . in Fig. 4 and the corresponding feature map is shown in Fig. 5.
382 Y. Lu et al. / Neurocomputing 267 (2017) 378–384

Table 2 chine [14] which achieves a better performance on object recog-

Recognition performance using mean-pooling, max-pooling and stochastic pooling.
nition. We can extend this model for fault diganosis. Moveover, we
Different pooling Mean-pooling Max-pooling Stochastic-pooling plan to have a deeper analysis of the training method with and
Recognition accuracy (%) 92.11 93.24 95.48 without labeled samples. The results in this paper could be further
extended to the distributed state estimation problems for sensor
networks [9] and nonlinear time-varying systems [8,11,13].
Table 3
Recognition performance using different convolutional filter size.

Different convolutional filter size 5×5 9×9 16 × 16 32 × 32 References

Recognition accuracy (% ) 93.15 92.56 93.29 92.48
[1] A. Asfarian, Y. Herdiyeni, A. Rauf, K.H. Mutaqin, A computer vision for rice
disease identification to support integrated pest management, Crop Prot. 61
Table 4 (2014) 103–104.
Simulation results using proposed method compared with [2] J.G.A. Barbedo, Digital image processing techniques for detecting, quantifying
and classifying plant diseases, SpringerPlus 2 (1) (2013) 660–672.
BP, SVM and PSO.
[3] Y. Bengio, Learning deep architectures for AI, Found. Trends Mach. Learn. 2 (1)
Algorithm CNN BP SVM PSO (2009) 1–127.
[4] J. Bouvrie, Notes on convolutional neural networks, Technical report, 2006.
Accuracy rate (%) 95 92 91 88 [5] Y. Chai, X.D. Wang, Recognition of greenhouse tomato disease based on image
Missing report rates (%) 5 8 9 12 processing technology, Tech. Autom. Appl. 9 (2013) 83–89.
False report rate (%) 0 0 5 10 [6] D.C. Ciresan, U. Meier, J. Masci, L.M. Gambardella, J. Schmidhuber, Flexible,
high performance convolutional neural networks for image classification, in:
Proceedings of the International Joint Conference on Artificial Intelligence (IJ-
CAI), Barcelona, Spain, 22, 2011, pp. 1238–1242.
Before training, we need to normalize the brightness and con- [7] F.R. Chen, Rice Diseases and Insect Pests Atlas and Control Technology, China
trast of all the training samples patches. The experiment is that, for Agriculture Science Press, Beijing, China, 2012.
[8] H. Dong, Z. Wang, S.X. Ding, H. Gao, Finite-horizon estimation of randomly
each sample, we subtract the mean and divide the standard devia- occurring faults for a class of nonlinear time-varying systems, Automatica 50
tion. In addition, when divided by the standard deviation with the (12) (2014) 3182–3189.
denominator being 0, in order to avoid and suppress noise, we give [9] H. Dong, Z. Wang, F.E. Alsaadi, B. Ahmad, Event-triggered robust distributed
state estimation for sensor networks with state-dependent noises, Int. J. Gen.
the standard deviation of adding a small constant. For the range of Syst. 44 (2) (2015) 254–266.
gray scale [0, 255], the variance plus 10 can get better results. [10] H. Dong, Z. Wang, S.X. Ding, H. Gao, Finite-horizon reliable control with ran-
Under the 10-fold cross-validation strategy to train and test, and domly occurring uncertainties and nonlinearities subject to output quantiza-
tion, Automatica 52 (2015) 355–362.
compare the mean-pooling, max-pooling and stochastic pooling, [11] H. Dong, Z. Wang, S.X. Ding, H. Gao, Event-based h∞ filter design for a class
the results are shown in Table 2, from which it can be seen that of nonlinear time-varying systems with fading channels and multiplicative
stochastic pooling gained better results. Also, compare the convo- noises, IEEE Trans. Signal Process. 63 (13) (2015) 3387–3395.
[12] H. Dong, Z. Wang, B. Shen, D. Ding, Variance-constrained h-infinity control for
lutional filter sizes of 5 × 5, 9 × 9, 16 × 16 and 32 × 32, the recogni-
a class of nonlinear stochastic discrete time-varying systems: the event-trig-
tion performance is shown in Table 3. gered design, Automatica 72 (2016) 28–36.
As we can see, different convolutional filter sizes have little ef- [13] H. Dong, Z. Wang, S.X. Ding, H. Gao, On h∞ estimation of randomly occurring
faults for a class of nonlinear time-varying systems with fading channels, IEEE
fect on the recognition results.
Trans. Autom. Control 61 (2) (2016) 479–484.
The compared results with BP method, support vector machine [14] G.E. Hinton, A practical guide to training restricted Boltzmann machines, Mo-
(SVM) method and particle swarm optimization (PSO) are shown mentum 9 (1) (2012) 599–619.
in Table 4. [15] S. Ji, W. Xu, M. Yang, K. Yu, 3D convolutional neural networks for human action
recognition, IEEE Trans. Pattern Anal. Mach. Intell. 35 (1) (2013) 221–231.
[16] Z. Jian, Z. Wei, Support vector machine for recognition of cucumber leaf dis-
5. Conclusions eases, in: Proceedings of the International Conference on Advanced Computer
Control (ICACC), Patiala, India, 2010, pp. 264–266.
[17] J. Jin, K. Fu, C. Zhang, Traffic sign recognition with hinge loss trained con-
CNNs is a valuable pattern-recognition method both in theory volutional neural networks, IEEE Trans. Intell. Transp. Syst. 15 (5) (2014)
and in application. In this paper, we proposed an innovative tech- 1991–20 0 0.
nique to enhance the deep learning ability of CNNs. The proposed [18] A. Karpathy, G. Toderici, S. Shetty, et al., Large-scale video classification with
convolutional neural networks, in: Proceedings of the IEEE Conference on
CNNs-based model can effectively classify 10 common rice diseases
Computer Vision and Pattern Recognition (CVPR), Columbus, OH, USA, 2014,
through images recognition. The application to the rice disease pp. 1725–1732.
identification shows that the proposed CNNs model can correctly [19] R. Kawasaki, H. Uga, S. Kagiwada, H. Iyatomi, Basic study of automated diagno-
and effectively recognize rice diseases through image recognition. sis of viral plant diseases using convolutional neural networks, in: Proceedings
of the International Symposium on Visual Computing (ISVC), Las Vegas, NV,
Compared with the other model, the proposed method has a better USA, 2015, pp. 638–645.
training performance, faster convergence rate, as well as a better [20] K. Khairnar, R. Dagade, Disease detection and diagnosis on plant using image
recognition ability than the other model. processing - a review, Int. J. Comput. Appl. 108 (13) (2014) 36–39.
[21] A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep con-
Deep convolution neural networks have achieved great perfor- volutional neural networks, Adv. Neural Inf. Process. Syst. (2012) 1097–1105.
mance breakthroughs in machine learning fields, but there still ex- Lake Tahoe, Nevada, USA
ist some research challenges. The first is that the CNNs may be [22] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (7553) (2015)
436–444.
composed of tens or hundreds of layers, each layer also includes [23] Y. LeCun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to
thousands of neurons, how many layers and how many neurons document recognition, Proc. IEEE 86 (11) (1998) 2278–2324.
are optimal at last? It has no choice but to rely on a great deal of [24] H. Lee, A. Battle, R. Raina, A.Y. Ng, Efficient sparse coding algorithms, Adv. Neu-
ral Inf. Process. Syst. (2006) 801–808.
experiments. The second is that the efficient deep learning algo-
[25] S. Phadikar, J. Sil, Rice disease identification using pattern recognition tech-
rithm is still depending on large-scale datasets, such as Imagenet niques, in: Proceedings of the IEEE International Conference on Computer and
ILSVRC containing millions of training examples. To improve rice Information Technology (ICCIT), Khulna, Bangladesh, 2008, pp. 420–423.
[26] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpa-
diseases identification accuracy, we still need to provide thousands
thy, A. Khosla, M. Bernstein, A.C. Berg, L. Fei-Fei, Imagenet large scale visual
of high-quality rice diseases images samples. Furthermore, there recognition challenge, Int. J. Comput. Vision 115 (3) (2015) 211–252.
are a large number of parameters in CNNs, and how to find the [27] R. Salakhutdinov, G.E. Hinton, Deep Boltzmann machines, in: Proceedings of
optimal parameters is also a research challenge. the International Conference on Artificial Intelligence and Statistics, Clearwater
Beach, Florida USA, 2009, pp. 448–455.
In future work, we plan to apply other deep architectures and [28] P. Sanyal, S.C. Patel, Pattern recognition method to detect two diseases in rice
other training algorithms, such as the restricted Boltzmann ma- plants, Imaging Sci. J. 56 (6) (2013) 319–325.
 Y. Lu et al. / Neurocomputing 267 (2017) 378–384 383

[29] Z. Shicha, M. Hanping, H. Bo, Z. Yancheng, Morphological feature extraction for [33] M. Zeiler, R. Fergus, Stochastic pooling for regularization of deep convolutional
cotton disease recognition by machine vision, Microcomput. Inf. 23 (4) (2007) neural networks, arXiv preprint (2013). arXiv:1301.3557.
290–292. [34] L.N. Zhang, B. Yang, Research on recognition of maize disease based on mo-
[30] D. Silver, A. Huang, C.J. Maddison, A. Guez, L. Sifre, et al., Mastering the game bile internet and support vector machine technique, Trans. Tech. Publ. 108 (13)
of go with deep neural networks and tree search, Nature 529 (7587) (2016) (2014) 659–662.
484–489. [35] X. Zhang, Y. Lu, S. Zhang, Multi-task learning for food identification and anal-
[31] K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale ysis with deep convolutional neural networks, J. Comput. Sci. Technol. 31 (3)
image recognition, arXiv preprint (2014). arXiv:1409.1556. (2016) 489–500.
[32] F.Y. Wang, Lets go: from AlphaGo to parallel intelligence, Sci. Technol. Rev. 34 [36] W.S. Zhao, Y.L. Peng, Rice Diseases and Insect Pests Atlas, China Agriculture
(7) (2016) 72–74. Press, Beijing, China, 2012.
384 Y. Lu et al. / Neurocomputing 267 (2017) 378–384

Yang Lu received his M.S. and Ph.D. degrees from North- Yurong Liu received his B.Sc. degree in Mathematics from
east Petroleum University in Computer Application in Suzhou University, Suzhou, China, in 1986, the M.Sc. de-
2005 and Oil and Gas Engineering in 2013, respectively. gree in Applied Mathematics from Nanjing University of
Now he is an associate professor at Heilongjiang Bayi Science and Technology, Nanjing, China, in 1989, and the
Agricultural University. His research interests include ma- Ph.D. degree in Applied Mathematics from Suzhou Uni-
chine learning and pattern recognition, computer vision versity, Suzhou, China, in 20 0 0. Dr. Liu is currently a pro-
and neural network. fessor in the Department of Mathematics at Yangzhou
University, China. He has published more than 50 pa-
pers in refereed international journals. His current in-
terests include neural networks, complex networks, non-
linear dynamics, time-delay systems, multiagent systems,
and chaotic dynamics.
Shujuan Yi received her M.S. degree in Heilongjiang Bayi
Yong Zhang received his M.S. degree in Signal and Sys-
Agricultural University in 1990 and the Ph.D. degree in
tem Processing from Harbin Engineering University in
Agricultural Mechanization and Automation in Northeast
2011. Now he is an associate professor at Northeast
Agricultural University in 2008. Now she is a professor
Petroleum University and he is currently pursuing the
at Heilongjiang Bayi Agricultural University. Her research
Ph.D. degree in Chemical Engineering and Technology in
interests include control of complex systems, automation
Northeast Petroleum University. His research interests in-
and agricultural mechanization.
clude deep learning, neural networks and intelligent in-
strument.

Nianyin Zeng was born in Fujian Province, China, in 1986.

He received the B.Eng. degree in electrical engineering
and automation in 2008 and the Ph.D. degree in elec-
trical engineering in 2013, both from Fuzhou University,
Fuzhou, China. From October 2012 to March 2013, he was
a research assistant in the Department of Electrical and
Electronic Engineering, the University of Hong Kong, Hong
Kong. Dr. Zeng is currently an assistant professor in the
Department of Instrumental and Electrical Engineering at
Xiamen University, China. He is the author or coauthor of
several technical papers and also a very active reviewer
for many international journals. His current research in-
terests include nonlinear filtering, time-series modeling,
and bioinformatics.