Content uploaded by Haiguang Wang
Author content
All content in this area was uploaded by Haiguang Wang on Dec 31, 2019
Content may be subject to copyright.
978-1-4577-2133-5/10/$26.00 ©2012 IEEE 253
2012 8th International Conference on Natural Computation (ICNC 2012)
Image Recognition of Plant Diseases Based on
Principal Component Analysis and Neural Networks
Haiguang Wang, Guanlin Li, Zhanhong Ma, Xiaolong Li
Department of Plant Pathology,
China Agricultural University,
Beijing 100193, China
Abstract—Plant disease identification based on image processing
could quickly and accurately provide useful information for the
prediction and control of plant diseases. In this study, 21 color
features, 4 shape features and 25 texture features were extracted
from the images of two kinds wheat diseases (wheat stripe rust
and wheat leaf rust) and two kinds of grape diseases (grape
downy mildew and grape powdery mildew), principal component
analysis (PCA) was performed for reducing dimensions in feature
data processing, and then neural networks including
backpropagation (BP) networks, radial basis function (RBF)
neural networks, generalized regression networks (GRNNs) and
probabilistic neural networks (PNNs) were used as the classifiers
to identify wheat diseases and grape diseases, respectively. The
results showed that these neural networks could be used for
image recognition of these diseases based on reducing dimensions
using PCA and acceptable fitting accuracies and prediction
accuracies could be obtained. For the two kinds of wheat
diseases, the optimal recognition result was obtained when image
recognition was conducted based on PCA and BP networks, and
the fitting accuracy and the prediction accuracy were both 100%.
For the two kinds of grape diseases, the optimal recognition
results were obtained when GRNNs and PNNs were used as the
classifiers after reducing the dimensions of feature data with
PCA, and the prediction accuracies were 94.29% with the fitting
accuracies equal to 100%.
Keywords- image recognition; plant diseases; principal
component analysis; neural networks
I. INTRODUCTION
There are many kinds of plant diseases in the world. The
diseases could cause quality decline of agricultural products
and serious yield losses, and even threaten food security.
Timely recognition and diagnosis of plant diseases is the basis
for taking disease control measures. Recognition and diagnosis
of plant diseases usually relies on the in-field visual
identification by the agricultural technicians. This approach
requires high professional knowledge and rich experience, and
needs many professional and technical personnel. Disease
diagnosis via pathogen detection requires satisfactory
laboratory conditions and more professional knowledge. Now
the pathogen detection methods based on the molecular
biological techniques were rapidly developed, and more
accurate results of disease diagnosis could be obtained by using
these methods. However, these methods need to be performed
by professional and technical personnel and also could not be
performed in the field. Moreover, these methods are time-
consuming and of high cost. Therefore, it is very necessary to
find out a simple and fast plant disease identification method
with high identification accuracy.
With the rapid development of information technology and
agricultural informatization, computer technology has played
an important role in acquisition, processing and communication
of plant disease information [1]. Now image recognition of
plant diseases is the widespread concern generated as a result
of the development of visual technologies and the
popularization of digital products. The studies on the
recognition and the automatic assessment of plant diseases
based on image processing have been reported [2], [3], [4], [5],
[6], [7], [8], [9], [10]. Computer automatic recognition and
diagnosis based on symptom images of plant diseases could
quickly and accurately provide disease information for
agricultural technicians and farmers and thus reduce the
dependence on agricultural technicians.
Image recognition of plant diseases is to extract the
characteristic feature information from the diseased regions in
the obtained images by using image processing techniques, and
then to achieve disease recognition by using pattern recognition
methods such as discriminant analysis [11], [12], neural
networks [9], [13], [14], [15], [16] and support vector machine
[17], [18], [19], [20], [21]. Generally, the extracted features
from the images of plant diseases include color features [12],
[19], [22], [23], shape features [24], texture features [25], and
so on. It is very important to extract the effective characteristic
features for the image recognition of plant diseases. However,
sometimes excessive features are extracted from the images for
the recognition of plant diseases. This would increase the
requirements for computer hardware and software and the
complexity of disease recognition, and greatly extend the
computation time. So it is necessary to reduce the dimensions
of the feature data while many features are used for disease
recognition based on image processing. The commonly used
methods to reduce the dimensions of the feature data include
principal component analysis (PCA) [26], [27], stepwise linear
regression method [12], and so on.
In order to find out a method for plant disease identification
based on image processing, image recognition of four kinds of
important plant diseases including wheat stripe rust caused by
Puccinia striiformis f. sp. tritici, wheat leaf rust caused by
Puccinia recondita f. sp. tritici, grape downy mildew caused by
This work was supported in part by Special Fund for Agro-scientific Research
in the Public Interest (200903004 and 200903035).
Corresponding author: Haiguang Wang, E-mail: wanghaiguang@cau.edu.cn.
254
Plasmopara uiticola and grape powdery mildew caused by
Uncinula necator, was conducted based on color features,
shape features and texture features extracted from the disease
images using PCA and neural networks including
backpropagation (BP) networks, radial basis function (RBF)
neural networks, generalized regression networks (GRNNs)
and probabilistic neural networks (PNNs).
II. MATERIALS AND METHODS
In this study, 185 digital images of plant diseases were
obtained by using common digital camera. The images were
divided into two groups according to the types of plants; one
group included 50 images of wheat stripe rust and 50 images of
wheat leaf rust, another group included 50 images of grape
downy mildew and 35 images of grape powdery mildew. The
size of the original plant disease images was 2592×1944 with
format of jpg, 24 bitmap. To improve the operation speed of
computer programs, the images of wheat diseases were cropped
to the size of 400 × 300, and the images of grape diseases were
compressed from 2592×1944 to 800×600 in the same
proportion without changing the image resolution using the
nearest neighbor interpolation method. Then the plant disease
images were denoised with median filter algorithm. After
images preprocessing, K_means clustering algorithm was used
to segment the plant disease images [28]. In MATALAB 7.6,
50 features including 21 color features, 4 shape features and 25
texture features were extracted from the segmented disease
images [29].
The image recognition of plant diseases was carried out
using BP networks, RBF neural networks, GRNNs and PNNs
as the classifiers, respectively. For wheat diseases, 30 images
of wheat stripe rust and 30 images of wheat leaf rust were
randomly selected as the training set, and the remaining wheat
disease images including 20 images of wheat stripe rust and 20
images of wheat leaf rust were regarded as the testing set. For
grape diseases, 30 images of grape downy mildew and 20
images of grape powdery mildew were randomly selected as
the training set, and the remaining grape disease images
including 20 images of grape downy mildew and 15 images of
grape powdery mildew were regarded as the testing set. For the
disease images of each kind of plant, seven groups of feature
combinations were obtained and were recorded as Col, Sha,
Tex, ColSha, ColTex, ShaTex and CST, respectively. Col
referred to 21 color features, Sha referred to 4 shape features,
Tex referred to 25 texture features, ColSha referred to 21 color
features and 4 shape features, ColTex referred to 21 color
features and 25 texture features, ShaTex referred to 4 shape
features and 25 texture features, and CST referred to 50
features. These seven groups of feature combinations were
processed by using PCA, respectively.
PCA is to convert a set of observations of possibly
correlated variables into a set of values of linearly uncorrelated
variables using an orthogonal transformation. Usually, it is
used for reducing dimensions in data processing. In this study,
the processes of PCA were implemented by using the
procedure PRINCOMP in the SAS System for Windows 8.2.
Principal component was selected for further processing if the
eigenvalue was about 1 and the cumulative contribution
rate≥85％. The selected principal components were normalized
and then were regarded as network inputs. The data on disease
types were regarded as target outputs. While BP networks,
RBF neural networks and GRNNs were used as the classifiers,
wheat stripe rust and wheat leaf rust were expressed as (0, 1)
and (1, 0), respectively, and grape downy mildew and grape
powdery mildew were expressed as (0, 1) and (1, 0) ,
respectively. While RBF neural networks, GRNNs and PNNs
were used as the classifiers, wheat stripe rust and wheat leaf
rust were expressed as 1 and 2, respectively, and grape downy
mildew and grape powdery mildew were expressed as 1 and 2,
respectively. The fitting recognition results and the prediction
recognition results were compared with the actual disease
types, and then the fitting accuracies and the prediction
accuracies were obtained, respectively.
In MATALAB 7.6, BP networks, RBF neural networks,
GRNNs and PNNs were designed with newff, newrbe,
newgrnn and newpnn, respectively. One-hidden-layer BP
networks were constructed for the image recognition of plant
diseases. The transfer function tansig was used in the hidden
layer and the log-sigmoid transfer function logsig was used in
the output layer. Levenberg-Marquardt algorithm (trainlm) was
used as training functions, and learngdm was used as learning
functions. For the BP networks, maximum number of epochs to
train was 5000 and the goal of training performance was 0.01.
Number of neurons in the hidden layer n was calculated using
the following formula,
annn ++= 21 (1)
in which，n1 is the number of neurons in the input layer, n2 is
the number of neurons in the output layer, and a is a constant
between 1 and 10. For RBF neural networks, GRNNs and
PNNs, spreads of radial basis functions were assumed to be 0.1
to 2.0 with step size 0.1.
III. RESULTS AND ANALYSIS
The recognition results of the four kinds of neural networks
with fitting accuracy ≥75% and prediction accuracy ≥75% were
shown in TABLE Ⅰ, TABLE Ⅱ, TABLE Ⅲ, TABLE Ⅳ, and
TABLE Ⅴ, respectively. After reducing the dimensions of the
data of the different feature combinations using PCA,
acceptable fitting accuracies and prediction accuracies for
image recognition could be obtained using the neural networks
as the classifiers. Overall, the recognition results based on the
combinations of different features were better than that based
on the individual features. The recognition results obtained by
using BP networks, GRNNs and PNNs were better than that
obtained by using RBF neural networks. Whether the types of
the plant diseases were expressed as (0, 1) and (1, 0) or
expressed as 1 and 2 when RBF neural networks or GRNNs
were used as the classifiers, the recognition results were not
affected. The recognition results for wheat diseases and grape
diseases obtained by using RBF neural networks were listed in
TABLE Ⅲ. When GRNNs and PNNs were used as the
classifiers for the image recognition of wheat diseases, there
were some differences between the recognition results of these
two kinds of methods, and the recognition results were shown
in TABLE Ⅱ with the note describing the differences.
However, the recognition results when GRNNs were used as
255
the classifiers for the image recognition of grape diseases were
the same as that when PNNs were used, and the corresponding
results were shown in the same table (TABLE Ⅴ).
When BP networks were used as the classifiers, the
recognition results for wheat diseases were shown in TABLE
Ⅰ. The fitting accuracy and the prediction accuracy were both
100% when CST was used as inputs with the number of
neurons in the hidden layer equal to 6. The fitting accuracy was
98.33% and the prediction accuracy was 100% when ShaTex
was used as inputs with the number of neurons in the hidden
layer equal to 11. When ShaTex was used as inputs with the
number of neurons in the hidden layer equal to 5 and CST was
used as inputs with the number of neurons in the hidden layer
equal to 9, the fitting accuracies were 100% and the prediction
accuracies were 97.50%. When CST or ColTex was used as
inputs with the number of neurons in the hidden layer equal to
8, the fitting accuracy was 100% and the prediction accuracy
was 95%.
As TABLE Ⅱ shown, with some exceptions, most of the
recognition results for wheat diseases obtained by using
GRNNs and PNNs were the same. The best recognition results
were obtained when Tex was used as inputs with the value of
spread equal to 0.3; the fitting accuracy was 96.67% when
GRNN was used, the fitting accuracy was 95% when PNN was
used, and the prediction accuracies obtained by using these two
kinds of neural networks were both 100%. Using GRNNs as
the classifiers based on Sha, the fitting accuracy was 90% and
the prediction accuracy was 95% when the value of spread was
equal to 1.0, the fitting accuracy was 88.33% and the prediction
accuracy was 95% when the value of spread was equal to 1.6,
and the fitting accuracy was 88.33% and the prediction
accuracy was 92.50% when the value of spread was equal to
1.9 or 2.0. Using PNNs as the classifiers based on Sha, the
fitting accuracy was 88.33% and the prediction accuracy was
95% when the value of spread was equal to 1.0, the fitting
accuracy was 88.33% and the prediction accuracy was 92.50%
when the value of spread was equal to 1.6, and the fitting
accuracy was 88.33% and the prediction accuracy was 95%
when the value of spread was equal to 1.9 or 2.0. When Tex
was used as inputs with the value equal to 1.6, the fitting
accuracy was 91.67% and the prediction accuracy was 97.50%
using GRNNs, however, the fitting accuracy was 91.67% and
the prediction accuracy was 100% using PNNs. The prediction
accuracies were 97.50% when ShaTex was used as inputs with
the values of spread equal to 0.3~0.7. The fitting accuracies
were 91.67% and the prediction accuracies were 100% when
ShaTex was used as inputs with the values of spread equal to
0.8~2.0. The fitting accuracies were 91.67% and the prediction
accuracies were 100% when Tex was used as inputs with the
values of spread equal to 0.6~1.5. The fitting accuracies were
91.67% and the prediction accuracies were 97.50% when Tex
was used as inputs with the values of spread equal to 0.4, 0.5
and 1.7~2.0.
The recognition results using RBF neural networks as the
classifiers were shown in TABLE Ⅲ. For wheat diseases,
there were the acceptable recognition results picked out when
ColSha, ColTex, ShaTex, Tex or CST was used as inputs. For
grape diseases, there were the acceptable recognition results
picked out only when ColTex or CST was used as inputs. For
wheat diseases, the fitting accuracy was 100% and the
prediction accuracy was 95% when ColTex was used as inputs
with the value of spread equal to 0.5, the fitting accuracy was
100% and the prediction accuracy was 92.50% when CST was
used as inputs with the value of spread equal to 0.5, and the
fitting accuracy was 100% and the prediction accuracy was
90% when ColTex was used as inputs with the value of spread
equal to 0.4 or when Tex was used as inputs with the value of
spread equal to 0.2. For grape diseases, the best prediction
accuracy was 80% with the fitting accuracy equal to 100%
when CST was used as inputs with the value of spread equal to
0.5.
When BP networks were used as the classifiers, the
recognition results for grape diseases were shown in TABLE
Ⅳ. The fitting accuracies were 100% and the prediction
TABLE I. FITTING RESULTS AND PREDICTION RESULTS FOR IMAGE RECOGNITION OF WHEAT DISEASES USING BP NETWORKS
Features Number of neurons in the
hidden layer
Fitting
accuracy
Prediction
accuracy Features Number of neurons in the
hidden layer
Fitting
accuracy
Prediction
accuracy
Col 5 98.33% 80% ShaTex 3, 9 98.33% 82.50%
Col 8 98.33% 82.50% ShaTex 10 98.33% 92.50%
Col 10 100% 77.50% ShaTex 11 98.33% 100%
ColSha 3 98.33% 87.50% ShaTex 6 100% 80%
ColSha 5, 11 100% 75% ShaTex 7 100% 90%
ColSha 7, 8 100% 77.50% ShaTex 5 100% 97.50%
ColSha 10 100% 82.50% Tex 3 98.33% 77.50%
ColSha 6, 9 100% 85% Tex 12 98.33% 82.50%
ColSha 4, 12 100% 87.50% Tex 4, 6 98.33% 85%
ColTex 4 98.33% 82.50% Tex 8, 10 98.33% 90%
ColTex 10 100% 82.50% Tex 7 100% 77.50%
ColTex 6, 9 100% 85% Tex 5 100% 87.50%
ColTex 3, 7, 11 100% 87.50% CST 3 95% 77.50%
ColTex 12 100% 90% CST 10 98.33% 77.50%
ColTex 8 100% 95% CST 7 98.33% 80%
Sha 3 95% 92.50% CST 4, 12 98.33% 92.50%
Sha 4, 5, 11 98.33% 75% CST 11 100% 82.50%
Sha 7, 9 98.33% 77.50% CST 5 100% 87.50%
Sha 6 98.33% 85% CST 8 100% 95%
Sha 8 98.33% 90% CST 9 100% 97.50%
ShaTex 12 96.67% 85% CST 6 100% 100%
256
accuracies were 91.43% when ColSha was used as inputs with
the number of neurons in the hidden layer equal to 4 and
ShaTex was used as inputs with the number of neurons in the
hidden layer equal to 5, 8 and 9. The fitting accuracy was 98%
and the prediction accuracy was 91.43% when ColSha was
used as inputs with the number of neurons in the hidden layer
equal to 3. In the remaining cases, the prediction accuracies
were less than 90%.
For grape diseases, the recognition results obtained by
using GRNNs and PNNs as the classifiers were shown in
TABLE Ⅴ. The best prediction accuracy was 94.29% with
the fitting accuracy equal to 100% when ColSha was used as
inputs with the value of spread equal to 0.2. The prediction
accuracies were less than 90% in the remaining cases.
TABLE II. FITTING RESULTS AND PREDICTION RESULTS FOR IMAGE RECOGNITION OF WHEAT DISEASES USING GRNNS AND PNNS
Features Spread Fitting
accuracy
Prediction
accuracy Features Spread Fitting
accuracy
Prediction
accuracy
Col 1.3~2.0 88.33% 82.50% Sha
0.3, 0.4, 0.5, 1.1~1.5, 1.6a, 1.0b, 1.9b,
2.0b 88.33% 95%
Col 1.2 90% 82.50% Sha 0.6, 0.7 90% 92.50%
Col 0.5~1.1 90% 85% Sha 0.2, 0.9, 1.0a 90% 95%
Col 0.4 90% 87.50% Sha 0.1 93.33% 95%
Col 0.3 91.67% 87.50% ShaTex 0.5, 0.6, 0.7 91.67% 97.50%
Col 0.2 96.67% 85% ShaTex 0.8~2.0 91.67% 100%
Col 0.1 100% 82.50% ShaTex 0.3, 0.4 95% 97.50%
ColSha 0.7~2.0 88.33% 85% ShaTex 0.2 96.67% 92.50%
ColSha 0.6 90% 85% ShaTex 0.1 100% 92.50%
ColSha 0.4, 0.5 91.67% 87.50% Tex 0.4, 0.5, 1.7~2.0, 1.6a 91.67% 97.50%
ColSha 0.3 95% 90% Tex 0.6~1.5, 1.6b 91.67% 100%
ColSha 0.2 98.33% 87.50% Tex 0.2 95% 95%
ColSha 0.1 100% 80% Tex 0.3 96.67%/95%c 100%
ColTex 0.5~2.0 93.33% 95% Tex 0.1 100% 92.50%
ColTex 0.4 96.67% 95% CST 0.5~2.0 93.33% 95%
ColTex 0.3 98.33% 92.50% CST 0.4 96.67% 95%
ColTex 0.1 100% 92.50% CST 0.3 98.33% 92.50%
ColTex 0.2 100% 95% CST 0.1 100% 90%
Sha 0.8, 1.7, 1.8, 1.9 a, 2.0a,
1.6b 88.33% 92.50% CST 0.2 100% 95%
a. The fitting accuracy and the prediction accuracy were obtained by using GRNN as the classifier. b. The fitting accuracy and the prediction accuracy were
obtained by using PNN as the classifier. c. The fitting accuracy was 96.67% by using GRNN as the classifier and the fitting accuracy was 95% by using PNN as
the classifier.
TABLE III. FITTING RESULTS AND PREDICTION RESULTS FOR IMAGE RECOGNITION OF WHEAT DISEASES AND GRAPE DISEASES USING RBF NEURAL NETWORKS
Features Spread Fitting accuracy Prediction accuracy Features Spread Fitting accuracy Prediction accuracy
ColShaa 0.2, 0.3 100% 77.50% Texa 0.2 100% 90%
ColTexa 0.8, 1.1~1.6 100% 75% CSTa 0.7 100% 75%
ColTexa 0.6, 0.7 100% 80% CSTa 0.4, 0.6 100% 87.50%
ColTexa 0.4 100% 90% CSTa 0.5 100% 92.50%
ColTexa 0.5 100% 95% ColTexb 0.5 100% 77.14%
ShaTexa 0.2 100% 77.50% CST b 0.6 100% 77.14%
ShaTexa 0.3 100% 82.50% CSTb 0.5 100% 80%
a. The results were for wheat diseases. b. The results were for grape diseases.
TABLE IV. FITTING RESULTS AND PREDICTION RESULTS FOR IMAGE RECOGNITION OF GRAPE DISEASES USING BP NETWORKS
Features Number of neurons in the
hidden layer
Fitting
accuracy
Prediction
accuracy Features Number of neurons in the
hidden layer
Fitting
accuracy
Prediction
accuracy
Col 5 98% 77.14% Sha 6, 7, 10 98% 85.71%
Col 6, 9 100% 77.14% Sha 12 100% 82.86%
Col 4 100% 80% ShaTex 3 96% 82.86%
ColSha 9 98% 82.86% ShaTex 4 98% 88.57%
ColSha 3 98% 91.43% ShaTex 7 100% 80%
ColSha 10, 12 100% 82.86% ShaTex 11, 12 100% 88.57%
ColSha 7 100% 85.71% ShaTex 5, 8, 9 100% 91.43%
ColSha 5 100% 88.57% Tex 3 92% 85.71%
ColSha 4 100% 91.43% Tex 4 94% 85.71%
ColTex 13 100% 77.14% Tex 5, 8 100% 88.57%
ColTex 11 100% 80% CST 8, 10 98% 80%
ColTex 10 100% 82.86% CST 13 100% 77.14%
Sha 3 86% 82.86% CST 5 100% 82.86%
Sha 5 90% 82.86% CST 11 100% 88.57%
Sha 9, 11 98% 82.86%
257
TABLE V. FITTING RESULTS AND PREDICTION RESULTS FOR IMAGE RECOGNITION OF GRAPE DISEASES USING GRNNS AND PNNS
Features Spread Fitting accuracy Prediction accuracy Features Spread Fitting accuracy Prediction accuracy
Col 0.3 86% 82.86% Sha 0.1 90% 85.71%
Col 0.2 92% 80% ShaTex 0.4 76% 80%
ColSha 0.3 90% 85.71% ShaTex 0.3 90% 85.71%
ColSha 0.1 100% 82.86% ShaTex 0.2 92% 85.71%
ColSha 0.2 100% 94.29% ShaTex 0.1 98% 82.86%
ColTex 0.6 76% 77.14% Tex 0.3 82% 80%
ColTex 0.5 88% 80% Tex 0.2 88% 80%
ColTex 0.4 94% 77.14% CST 0.5 86% 77.14%
ColTex 0.3 96% 77.14% CST 0.4 96% 80%
ColTex 0.1 100% 82.86% CST 0.3 98% 82.86%
ColTex 0.2 100% 85.71% CST 0.1 100% 80%
Sha 0.3 80% 77.14% CST 0.2 100% 88.57%
Sha 0.2 84% 85.71%
IV. CONCLUSIONS AND DISCUSSIO N
In this study, PCA and neural networks were used to
implement the image recognition of plant diseases based on the
extracted color features, shape features and texture features
from disease images and their combined features. As the results
showed, reducing the dimensions of the feature data extracted
from the images of plant diseases could reduce the running
time of the neural networks and acceptable recognition results
could be obtained. The method used in this study could also be
used for the image recognition of other plant diseases. In the
practical application, PCA could be used to reduce the
dimensions of the data extracted from the plant disease images
and then optimal neural networks could be constructed for
plant disease identification.
The optimal results for wheat diseases and grape diseases
obtained by using SVM based on the same images used in this
study were reported by Li [29]. The optimal result for wheat
stripe rust and wheat leaf rust was that the fitting accuracy was
96.67% and the prediction accuracy was 100%, and the optimal
result for grape downy mildew and grape powdery mildew was
that the fitting accuracy was 100% and the prediction accuracy
was 91.43% [29]. In this study, for the two kinds of wheat
diseases, the optimal result based on PCA and BP networks
was that the fitting accuracy and the prediction accuracy were
both 100%, that based on PCA and GRNNs was the same as
that obtained by Li, that based on PCA and RBF neural
networks was that the fitting accuracy was 100% and the
prediction accuracy was 95%, and that based on PCA and
PNNs was that the fitting accuracy was 95% and the prediction
accuracy was 100%. In this study, for the two kinds of grape
diseases, the optimal result based on PCA and BP networks
was the same as that obtained by Li, that based on PCA and
GRNNs was that the fitting accuracy was 100% and the
prediction accuracy was 94.29%, that based on PCA and PNNs
was the same as that based on PCA and GRNNs, and that based
on PCA and RBF neural networks was that the fitting accuracy
was 100% and the prediction accuracy was 80%.
PCA could reduce the dimensions of the obtained data
under the premise of retaining the total data information,
reduce the number of neurons in the input layer and increase
the speed of the neural networks. PCA is a kind of linear
projection and it could not correctly handle non-linear data. In
the image recognition of plant diseases, some other methods
should be used to reduce dimensions of the feature data if the
extracted features are on-linear data.
If the image recognition of plant diseases is carried out on
personal computers, it is not necessary to perform the operation
of reducing the dimensions of the obtained data and the
recognition speed could not be affected significantly because of
the increasing computer performance and the strong ability of
the neural networks to solve problems. When image
recognition of plant diseases based on many extracted features
is carried out via the internet, it is better to reduce the
dimensions of the obtained data firstly and then to conduct
image recognition using neural networks.
REFERENCES
[1] H. G. Wang, Z. H. Ma, M. R. Zhang, and S. D. Shi, “Application of
Computer Technology in Plant Pathology”, Agriculture Network
Information. vol. 19, pp. 31–34, October 2004.
[2] C. C. Tucker and S. Chakraborty, “Quantitative assessment of lesion
characteristics and disease severity using digital image processing”, J.
Phytopathology, vol. 145, pp. 273–278, July 1997.
[3] I. S. Ahmad, J. F. Reid, M. R. Paulsen, and J. B. Sinclair, “Color
classifier for symptomatic soybean seeds using image processing”, Plant
Disease, vol. 83, pp. 320–327, April 1999.
[4] J. W. Olmstead, A. Gregory, and G. A. Lang, “Assessment of severity of
powdery mildew infection of sweet cherry leaves by digital image
analysis”, Hortscience, vol. 36, pp. 107–111, January 2001.
[5] J. C. Lai, S. K. Li, B. Ming, N. Wang, K. R. Wang, R. Z. Xie, and S. J.
Gao, “Advances in research on computer-vision diagnosis of crop
diseases”, Scientia Agricultura Sinica, vol. 42, pp. 1215–1221, April
2009.
[6] Z. X. Guan, Q. Yao, B. J. Yang, J. Hu, and J. Tang, “Application of
digital image processing technology in recognizing the diseases, pests,
and weeds from crops”, Scientia Agricultura Sinica, vol. 42, pp. 2349–
2358, July 2009.
[7] C. H. Bock, A. Z. Cook, P. E. Parker, and T. R. Gottwald, “Automated
image analysis of the severity of foliar citrus canker symptoms”, Plant
Disease, vol. 93, pp. 660–665, June 2009.
[8] C. H. Bock, G. H. Poole, P. E. Parker, and T. R. Gottwald, “Plant
disease severity estimated visually, by digital photography and image
analysis, and by hyperspectral imaging”, Critical Reviews in Plant
Sciences, vol. 29, pp. 59–107, March 2010.
[9] H. Al-Hiary, S. Bani-Ahmad, M. Reyalat, M. Braik, and Z.
ALRahamneh, “Fast and accurate detection and classification of plant
diseases”, International Journal of Computer Applications, vol. 17, pp.
31–38, March 2011.
[10] G. L. Li, Z. H. Ma, and H. G. Wang, “An automatic grading method of
severity of single leaf infected with grape downy mildew based on
image processing”, Journal of China Agricultural University, vol. 16, pp.
88–93, December 2011.
[11] N. Wang, K. R. Wang, R. Z. Xie, J. C. Lai, B. Ming, and S. K. Li,
“Maize leaf disease identification based on Fisher discrimination
analysis”, Scientia Agricultura Sinica, vol. 42, pp. 3836–3842,
November 2009.
258
[12] Z. X. Cen, B. J. Li, Y. X. Shi, H. Y. Huang, J. Liu, N. F. Liao, J. Feng,
“Discrimination of cucumber anthracnose and cucumber brown speck
base on color image statistical characteristics”, Acta Horticulturae
Sinica, vol. 34, pp. 1425–1430, June 2007.
[13] D. W. Zhang and J. Wang, “Design on image features recognition
system of cucumber downy mildew based on BP algorithm”, Journal of
Shenyang Jianzhu University (Natural Science), vol. 25, pp. 574–578,
May 2009.
[14] L. B. Liu and G. M. Zhou, “Identification method of rice leaf blast using
multilayer perception neural network”, Transactions of the CSAE, vol.
25(Supp.2), pp. 213–217, October 2009.
[15] Z. R. Li and D. J. He, “Research on identify technologies of apple’s
disease based on mobile photograph image analysis”, Computer
Engineering and Design, vol. 31, pp. 3051–3053, 3095, July 2010.
[16] D. T. Zhao, Y. H. Chai, and C. L. Zhang, “Inspection of soybean frogeye
spot based on image procession”, Journal of Northeast Agricultural
University, vol. 41, pp. 119–124, April 2010.
[17] Y. W. Tian, T. L. Li, C. H. Li, Z. L. Piao, G. K. Sun, and B. Wang,
“Method for recognition of grape disease based on support vector
machine”, Transactions of the CSAE, vol. 23, pp. 175–180, June 2007.
[18] D. Ren, H.Y. Yu, and J. H. Wang, “Research on plant disease
recognition based on linear combination of the kernel function support
vector machine”, Journal of Agricultural Mechanization Research, vol.
29, pp. 41–43, September 2007.
[19] Y. W. Tian, and Y. Niu, “Applied research of support vector machine on
recognition of cucumber disease”, Journal of Agricultural Mechanization
Research, vol. 31, pp. 36–39, March 2009.
[20] A. Camargo and J. S. Smith, “Image pattern classification for the
identification of disease causing agents in plants”, Computers and
Electronics in Agriculture, vol. 66, pp. 121–125, February 2009.
[21] G. L. Li, Z. H. Ma, and H. G. Wang, “Image recognition of grape downy
mildew and grape powdery mildew based on support vector machine”,
CCTA 2011, Part III, IFIP AICT 370, 2012, pp. 151–162.
[22] Y. L. Cui, P. F. Cheng, X. Z. Dong, Z. H. Liu, and S. X. Wang, “Image
processing and extracting color features of greenhouse diseased leaf”,
Transactions of the CSAE. vol. 21(Supp.), pp. 32–35, December 2005.
[23] P. Sanyala, and S. C. Patel, “Pattern recognition method to detect two
diseases in rice plants”, Imaging Science Journal, vol. 56, pp. 319–325,
December 2008.
[24] Y. X. Zhao, K. R. Wang, Z. Y. Bai, S. K. Li, R. Z. Xie, and S. J. Gao,
“Research of maize leaf disease identifying system based image
recognition”, Scientia Agricultura Sinica, vol. 40, pp. 698–703, April
2007.
[25] R. Pydipati, T. F. Burks, and W. S. Lee, “Identification of citrus disease
using color texture features and discriminant analysis”, Computers and
Electronics in Agriculture, vol. 52, pp. 49–59, June 2006.
[26] Z. Y. Liu, J. F. Huang, R. X. Tao, and H. Z. Zhang, “Estimating rice
brown spot disease severity based on principal component analysis and
radial basis function neural network”, Spectroscopy and Spectral
Analysis, vol. 28, pp. 2156–2160, September 2008.
[27] B. Li, Z. Y. Liu, J. F. Huang, L. L. Zhang, W. Zhou, and J. J. Shi,
“Hyperspectral identification of rice diseases and pests based on
principal component analysis and probabilistic neural network”,
Transactions of the CSAE, vol. 25, pp. 143–147, September 2009.
[28] G. L. Li, Z. H. Ma, C. Huang, Y. W. Chi, and H. G. Wang,
“Segmentation of color images of grape diseases using K_means
clustering algorithm”, Transactions of the CSAE, vol. 26(Supp.2), pp.
32–37, December 2010.
[29] G. L. Li, “Preliminary study on automatic diagnosis and classification
method of plant diseases based on image recognition technology”,
Beijing: China Agricultural University, 2011, pp. 1–64.
