Plant Diseases Recognition from Digital Images using Multichannel

Convolutional Neural Networks

Andre da Silva Abade

, Ana Paula G. S. de Almeida

and Flavio de Barros Vidal

Federal Institute of Education, Science and Technology of Mato Grosso, Brazil

Department of Mechanical Engineering, University of Brasilia, Distrito Federal, Brazil

Department of Computer Science, University of Brasilia, Distrito Federal, Brazil

Keywords:

Convolutional Neural Networks, Multichannel Convolutional Neural Networks, Plant Disease, Crop Disease

Recognition, Computer Vision.

Abstract:

Plant diseases are considered one of the main factors inﬂuencing food production and to minimize losses in

production, it is essential that crop diseases have a fast detection and recognition. Nowadays, recent studies use

deep learning techniques to diagnose plant diseases in an attempt to solve the main problem: a fast, low-cost

and efﬁcient methodology to diagnose plant diseases. In this work, we propose the use of classical convoluti-

onal neural network (CNN) models trained from scratch and a Multichannel CNN (M-CNN) approach to train

and evaluate the PlantVillage dataset, containing several plant diseases and more than 54,000 images (divided

into 38 diseases classes with 14 plant species). In both proposed approaches, our results achieved better accu-

racies than the state-of-the-art, with faster convergence and without the use of transfer learning techniques.

Our multichannel approach also demonstrates that the three versions of the dataset (colored, grayscaled and

segmented) can contribute to improve accuracy, adding relevant information to the proposed artiﬁcial neural

network.

1 INTRODUCTION

Plant diseases are considered one of the main factors

inﬂuencing food production, being responsible for the

signiﬁcant reduction of the physical or economic pro-

ductivity of the crops and, in certain cases, may be

an impediment to this activity. According to Altieri

(2018), in order to minimize production losses and

maintain crop sustainability, it is essential that dise-

ase management and control measures be carried out

in an appropriate manner, highlighting the constant

monitoring of the crop, combined with the rapid and

accurate diagnosis of the diseases. These practices are

the most recommended by phytopathologists.

The major challenge for agriculture is the correct

identiﬁcation of the symptoms of major diseases that

affect crops (Anderson et al., 2004). Manual and me-

chanized practices in traditional planting processes

are not able to cover large areas of plantation and pro-

vide essential early information to decision-making

processes (Miller et al., 2009). Thus, it is necessary

to develop automated solutions, practical, reliable and

economical able to monitor the health of plants provi-

ding meaningful information to the decision-making

process (e.g. correct dosage of pesticides (Mahlein,

2016)).

Computer Vision along with Artiﬁcial Intelligence

(AI) has been developing techniques and methods

for recognizing and classifying objects with signiﬁ-

cant advances (Arnal Barbedo, 2013). These systems

use Convolutional Neural Networks (CNNs) (Lecun

et al., 1998) and their results in some experiments

are already superior to humans in large-scale recon-

naissance tasks. The studies presented in Mohanty

et al. (2016) and Ferentinos (2018) make use of deep

learning techniques in agriculture, in particular in

the diagnosis of plant diseases. These approaches

have used two popular architectures, namely Alex-

Net (Krizhevsky et al., 2012) and GoogLeNet (Incep-

tion v1) (Szegedy et al., 2014), which were designed

in the context of the Scale Visual Recognition Chal-

lenge (ILSVRC) (Russakovsky et al., 2014) for the

ImageNet dataset (Deng et al., 2009).

With the aforementioned architectures, Mohanty

et al. (2016) show that only the colored dataset is suf-

ﬁcient to recognize plant diseases. However, more in-

formation about a subject can contribute to improve

the network accuracy and to conﬁrm this assump-

450

Abade, A., S. de Almeida, A. and Vidal, F.

Plant Diseases Recognition from Digital Images using Multichannel Convolutional Neural Networks.

DOI: 10.5220/0007383904500458

In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 450-458

ISBN: 978-989-758-354-4

tion, other versions – gray-scale and segmented – of

the PlantVillage (Hughes et al., 2015) dataset were

combined into multichannel convolutional neural net-

works, using the same architectures to our models for

a fair comparison. Furthermore, our work also im-

proves the reference single channel baseline without

using transfer learning techniques.

The paper is organized as follows: In Section 2

are presented the related works about plant disease

recognition. Section 3 explains our methodology.

Section 4 contains information about the results and

discussion. Section 5 provides conclusion points and

further works.

2 RELATED WORDS

The recognition and classiﬁcation of leaf diseases of

plants is a problem with many challenges to over-

come. The analysis in the process of identiﬁcation

of the diseases through the leaves can incur a large

number of false positives, for example, the symptoms

of phytotoxicity are associated with some disease due

to similar leaf lesions.

We developed an extensive time-review (from

most to less recent) of the main literature works, from

the traditional techniques and methods used in the

process of recognition and classiﬁcation of foliar di-

seases in plants to the latest advances provided by the

use of Convolutional Neural Networks (CNNs), sin-

gle and multichannel approaches. Table 1 presents

these works in chronological order, summarizing the

used techniques and methods and their consequent

area of application.

Before the advent of CNNs, traditional machine

learning classiﬁcation methods, such as SVM (Rumpf

et al., 2010) and K-Means (Al-hiary et al., 2011),

were used to classify diseases in plants. Patil

and Bodhe (2011) applied classic image processing

technique for disease detection in sugarcane leaves by

using threshold segmentation to determine leaf area

and triangle threshold for lesioning area, getting the

average accuracy of 98.60%. An approach proposed

by Singh and Misra (2017) uses genetic algorithms

for image segmentation which is an important aspect

for disease detection in a plant leaf.

Relevant works approach feature extraction

techniques for the detection of plant diseases. It is

possible to highlight the studies of Pydipati et al.

(2006), where there is use of color co-occurrence

method (CCM) to determine whether texture based

hue, saturation, and intensity (HSI) color features in

conjunction with statistical classiﬁcation algorithms

could be used to identify diseased and normal citrus

leaves under laboratory conditions. The leaf sam-

ple discriminant analysis using CCM textural featu-

res achieved classiﬁcation accuracies of over 95% for

all classes when using hue and saturation texture fe-

atures. According to Jabal et al. (2013), feature ex-

traction is a promising approach capable of solving

dichotomies between datasets constructed with ima-

ges in controlled environments and images captured

in the real world. This study proposed an ideal case

approach in plant classiﬁcation and recognition that

was not only applicable in the real world, but also

acceptable in laboratory conditions.

Due to the increase in processing capacity trig-

gered by the use of Graphics Process Unit (GPU),

AI is corroborating signiﬁcantly with the robust set

of traditional resources applied by Computer Vision

techniques (Ferentinos, 2018). Tacitly, Machine Le-

arning techniques have demonstrated signiﬁcant gains

in accuracy in the process of classiﬁcation and identi-

ﬁcation of plant diseases.

These advances are demonstrated in the works

of Rumpf et al. (2010), which proposes an approach

for the detection and differentiation of plant disea-

ses using Support Vector Machine algorithms. In this

study, the authors implemented a technique to iden-

tify beet diseases, in which depending on the type

and stage of disease the classiﬁcation accuracy was

between 65% and 90%. Another approach based on

leaf images and using Artiﬁcial Neural Networks as

a technique for an automatic detection and classiﬁca-

tion of plant diseases was used in conjunction with

K-means as a clustering procedure proposed in the

works of Al-hiary et al. (2011). On average, the accu-

racy of classiﬁcation using this approach was 94.67%.

According to LeCun et al. (2015), deep learning

allows computational models to learn representations

of data with multiple levels of abstraction, improving

the state-of-the-art in many domains, such as speech

recognition, object recognition, object detection. One

particular type of deep, feedforward network that was

much easier to train and generalized much better than

networks with full connectivity was the convolutio-

nal neural networks (CNNs). CNNs constitute one

of the most powerful techniques for modeling com-

plex processes and performing pattern recognition in

applications with large amount of data, like the one

of pattern recognition in images (LeCun et al., 2015).

Sladojevic et al. (2016) develops a model using CNN

capable of recognizing 13 different types of diseases

of healthy leafy plants, with the ability to distinguish

the leaves of the plants from their surroundings. The

experimental results on the developed model achieved

precision between 91% and 98%, for separate class

tests, on average 96.3%.

Plant Diseases Recognition from Digital Images using Multichannel Convolutional Neural Networks

451

Table 1: Review on the methods and techniques of leaf plant diseases’ recognition and classiﬁcation.

Year Author Method Application area

2018 Ferrentinos Convolutional neural network Identiﬁcation of leaf disease from 25 different species of plants

2018 Lin Zhongqi et al. Multichannel Convolutional neural network Detecting maize leaf diseases for 5 diseases

2017 Yang Lu et al. Convolutional neural network Identiﬁcation of rice diseases

2017 Pawara et al. Local descriptors and CNN Identiﬁcation of fruits diseases

2017 Tallha Akram et al. Based on an Image processing technique Real time classiﬁcation of plant diseases

2017 Trimi Neha Tete et al. Neural network, K-means and thresholding Identiﬁcation of disease from potato, apple and mango leaves

2017 Vijai Singh et al. Image segmentation technique Detection of plant leaf diseases

2017 Megha S. et al. Fuzzy c means and Support vector machine Identiﬁcation of plant leaf disease

2017 Lin Yuan et al. Fisher’s linear discriminant analysis (FLDA) Identiﬁcation of plant diseases and pests form SAR images

2016 Mohanty et al. Convolutional neural network Identiﬁcation of leaf disease from 25 different species of plants

2016 Sladojevic et al. Convolutional neural network Identiﬁcation of plant leaf disease

2016 Pujari et al.

Support vector machine and

Artiﬁcial neural network

Identiﬁcation of plant leaf disease of crops such as wheat,

maize, grape, sunﬂower etc.

2016 Ramakrishnan M. et al. Backpropagation algorithm Identiﬁcation of groundnut leaf disease

2016 Malvika Ranjan et al. Artiﬁcial neural network Identiﬁcation of cotton leaf disease

2015 Praksh M. Mainkar et al.

K-means clustering, GLCM and

Backpropagation neural network

Identiﬁcation of disease from potato, tomato and cotton leaves

2014 Marion Neumann et al. Support vector machine Identiﬁcation of beet leaf disease

2014 Rong Zhou et al. Support vector machine Identiﬁcation of Cercospora Leaf Spot from Sugar beet

2013 Jabal et al. Features extraction Recognition and classiﬁcation of plant leaf disease

2011 Patil et al. Based on an Image processing technique Identiﬁcation of plant leaf disease

2011 Al-hiary et al. K-menas clustering Identiﬁcation of plant leaf disease

2010 T. Rumpf et al. Support vector machine Identiﬁcation of Sugar beet disease from leaves

2006 Pydipati et al. Color texture features Identiﬁcation of Citrus disease

Using a public dataset of 54,306 images of disea-

sed and healthy plant leaves collected under control-

led conditions, the study of Mohanty et al. (2016)

train a deep neural network to identify 14 crop spe-

cies and 26 diseases (or absence thereof). The trained

model achieves an accuracy of 99.35% on a held-out

test set, demonstrating the feasibility of this appro-

ach. Pawara et al. (2017) compared the performance

of some conventional pattern recognition techniques

with that of CNN models, in plants identiﬁcation,

using three different databases of images of either en-

tire plants and fruits, or plant leaves, concluding that

CNNs drastically outperform conventional methods.

The propose described in Lu et al. (2017) pre-

sents a novel rice diseases identiﬁcation method ba-

sed on CNN techniques. Using a dataset of 500 natu-

ral images of diseased and healthy rice leaves, CNNs

are trained to identify 10 common rice diseases. Un-

der the 10-fold cross-validation strategy, the proposed

CNNs-based model achieves an accuracy of 95.48%.

Finally, Ferentinos (2018) used CNN models with an

open database of 87,848 images, containing 25 dif-

ferent plants in a set of 58 distinct classes of [plant,

disease] combinations, including healthy plants. Se-

veral model architectures were trained, with the best

performance reaching a 99.53% success rate in identi-

fying the corresponding [plant, disease] combination

(or healthy plant).

In the context of a M-CNN approach, Lin et al.

(2018) describes a simple use of the M-CNN archi-

tecture to detect and recognize maize leaf diseases,

using a dataset of 10,820 images containing ﬁve com-

mon maize leaf diseases. This approach uses a Re-

gion of Interest (ROI) to preprocess the input image

and achieves an accuracy of 92.31% with 30,000 ite-

rations/epochs. Even though this result was not better

than all of the single channel CNNs approaches des-

cribed earlier in this Section, the use of a reduced da-

taset in Lin et al. (2018) indicates that a M-CNN may

be a relevant and improved approach for plant disea-

ses detection and recognition.

3 PROPOSED METHODOLOGY

The research process of this study was guided by the

work of Mohanty et al. (2016). The state-of-the-art

shown by the author motivated our efforts to improve

not only the accuracy achieved by the methods pre-

viously proposed, but also develop and implement an

approach to produce more consistent results. In Mo-

hanty et al. (2016) work, it is shown that the colored

dataset is sufﬁcient to perform the recognition of plant

diseases. Our work combines in a multichannel con-

volutional neural network (M-CNN) the other availa-

ble versions of the dataset in the same model in or-

der to improve the network accuracy. Also, our work

improves the single channel CNN’s baseline without

using transfer learning techniques. The chosen trai-

ning/testing ratio was 80/20, the ratio that produced

better results in the reference work.

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

452

3.1 Dataset

The proposed methodology uses the PlantVillage da-

taset, provided by Hughes et al. (2015), containing

54,306 images of plant leaves and 38 different classes,

each class corresponding to a different crop disease.

Every class has three different versions: original co-

lored image, grayscaled image and segmented image.

Figures 1-(A) to (C) show a sample of this dataset.

Figure 1: Examples of plant diseases Dataset: (A) The three

versions of apple scab disease. (B) A sample of black rot, a

grape disease. (C) Three versions of strawberry leaf scorch.

3.2 CNN Architecture

According to Goodfellow et al. (2016), CNNs are spe-

cialized artiﬁcial neural networks that process input

data with some kind of spatial topology, such as ima-

ges, videos, audio and text. In addition to convo-

lution layers, CNNs are usually composed of other

types of layers, such as pooling. In this work, two

classic CNNs are used to evaluate the dataset: Alex-

Net (Krizhevsky et al., 2012) and GoogLeNet (Incep-

tion v1) (Szegedy et al., 2014).

3.2.1 Multichannel CNNs

Multichannel CNNs (M-CNNs) are generally used

when parallel processing of the input data is desi-

red (Karpathy et al., 2014). Such streams can eventu-

ally merge into one in the latter layers of the network.

In the relevant contributions of the studies of Baccou-

che et al. (2011) and Ji et al. (2013), it is common

for the point of concatenation to be present before the

ﬁrst fully connected layer of the network, that is, the

parallel processing is concentrated between the con-

volution layers. In Karpathy et al. (2014), a 2-channel

CNN is proposed, each channel receiving two frames

of the input video and being capable of generating la-

bels of the main action. Another advantage of using

M-CNNs is also highlighted by Karpathy et al. (2014)

and it consists in reducing the dimensionality of the

network input, which helps to decrease the proces-

sing time. In Figure 2 it is presented a generic archi-

tecture of a network with two input channels. Each

channel receives one different type of the dataset, ge-

nerating three additional versions:Version 1: Color +

Grayscale; Version 2: Color + Segmented; Version

3: Grayscale + Segmented.

There were not any pre-processing steps and all

the images had the same resolution size of 256 × 256

pixels. The objective behind the use of multichannel

networks is to observe whether the neural network can

produce better results if additional information is pro-

vided. Our models use a late fusion technique (Kar-

pathy et al., 2014), where two separate single channel

networks with shared parameters are merged in the

ﬁrst fully connected layer, computing global features

by comparing outputs of both streams.

To improve the single channel baseline, a hyperpa-

rameter optimization strategy was used. The optimi-

zation of the hyperparameters in the training of CNNs

is a process that demands a lot of effort, due to the nu-

merous parameters that can be adjusted, to the context

of the input data, to the deep learning network model

used and the deﬁned architecture. In this study we

adopted the grid search capability to adjust the hyper-

parameters of each learning model. Our hyperpara-

meter optimization strategy makes use of a reference

value as the starting point for exploring a range of va-

lues according to each parameter that can be adjusted.

The grid search was constructed with reference to the

values presented in study of Mohanty et al. (2016).

Thus, speciﬁc values were selected to achieve the best

results during training.

According to Srivastava et al. (2014), dropout

is a regularization technique for reducing overﬁt-

ting in neural networks by preventing complex co-

adaptations on training data.

Basically, half of the neurons on a particular layer

will be deactivated during training. The generaliza-

tion is improved due to the forcing of your layer to

learn the same ”concept” with different neurons. Nor-

mally, some deep learning models use dropout on the

fully connected layers, but is also possible to use dro-

pout after the max-pooling layers, creating some kind

of image noise augmentation.

In our approach, dropout layers are added before

and after the fusion that occurs in M-CNN architec-

Plant Diseases Recognition from Digital Images using Multichannel Convolutional Neural Networks

453

Figure 2: An illustration of the generic architecture of a multichannel convolutional network. The model generalizes a

structure with two input channels and identiﬁes the most relevant segments of architecture.

ture. We insert a dropout layer between the pooling

layers that precede the fusion of the networks and af-

ter the fusion, for the ﬁrst two fully connected layers.

In addition, based on the observations about the va-

lues of loss, we adjust the fraction of inputs to 0 at

each update during training time, which helps prevent

overﬁtting.

All CNN models (single and multichannel) were

trained using the training parameters presented in Ta-

ble 2.

Transfer learning is the technique of training

a base network on a base dataset, usually Image-

Net (Deng et al., 2009), and then transfer the learned

features to a second target network to be trained on

a target dataset and task (Yosinski et al., 2014). Fre-

quently, this approach tends to improve network over-

all accuracy, as seen in the results of Mohanty et al.

(2016). One of our goals in this work was to out-

perform the state-of-the-art values without the use of

transfer learning.

According to Pan and Yang (2010), transfer lear-

ning techniques are advantageous when used in CNNs

because they shorten training time since initial weig-

hts are imported from a similar training experience

performed on a larger data set. Thus, it is possible

to increase the accuracy of a CNN using transfer le-

arning even though its dataset is noticeably smaller.

However, it should be noted that if the dataset fea-

tures to be used has unique peculiarities to its set of

objects and the input images have different dimensi-

ons of pre-trained model, the use of learning transfer

technique should be rethought.

In the development of this study, after analyzing

the data set with its relevant characteristics, the mo-

del and architecture of each convolutional neural net-

work used, we chose starting the learning process

from scratch.

Table 2: Training hyperparameters.

Model Hyperparameters Values

AlexNet

Learning rate 0.01

Momentum 0.9

Weight decay 1e-6

Batch size 128

GoogleNet

Learning rate 0.0205

Momentum 0.9

Weight decay 0.0005

Batch size 16

4 RESULTS

This section presents the results obtained based on

the CNN architectures detailed in the methodology.

The GPU used for training the proposed models was a

NVIDIA GeForce GTX Titan Xp and all models were

developed using TensorFlow API version 1.6 (Abadi

et al., 2015) and Keras version 2.2.1 (Chollet et al.,

2015) frameworks. For evaluation, we used mean F

score and overall accuracy.

For better visualization, AlexNet and GoogleNet

with multichannel architecture will be named as M-

AlexNet and M-GoogleNet.

Table 3 shows the results of the proposed metho-

dology for single and multichannel architectures after

computing the mean F

score of each network, with

the best achieved result highlighted. In addition, we

present the best results obtained in the work of Mo-

hanty et al. (2016) as a reference for discussions. All

multichannel models were trained from scratch.

Figures 3 to 10 show the performance and losses

of all the testing models during the training process.

In Section 3, we report that we would use the trai-

ning process from scratch to compute the results. In

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

454

Table 3: Mean F

score of the proposed architectures and a comparison.

Model Dataset Type Mean F

Score

Mohanty et al. (2016)

AlexNet (transfer learning) Color 0.9927

GoogleNet (transfer learning) Color 0.9934

Ours

AlexNet (from scratch) Color 0.9873

GoogleNet (from scratch) Color 0.9940

M-AlexNet (from scratch)

Version 1 0.9959

Version 2 0.9920

Version 3 0.9923

M-GoogleNet (from scratch)

Version 1 0.9955

Version 2 0.9938

Version 3 0.9941

Table 3, our results are presented with signiﬁcant im-

provements, even though we do not use transfer lear-

ning techniques.

Basically, it is possible to explain this impro-

vement of results for the single channel from scratch

networks in comparison to the same proposal of the

work of Mohanty et al. (2016), by the singularity of

the characteristics of the dataset. Although the dataset

is considered small by its number of images, the mo-

dels pre-trained with the imaging do not have a signi-

ﬁcant sample space of labeled diseased plants. There-

fore, although the training time increased subtly, the

accuracy gains were representative.

In the study of Mohanty et al. (2016) its con-

clusions make it understood that the grayscaled and

segmented versions do not collaborate for an impro-

vement of the accuracy when comparing to the co-

lored version of the dataset. In our evaluation after

the experiments, it was possible to observe that for

models with single channel architecture that assump-

tion remains consistent, even though they were trai-

ned from scratch and with hyperparameters optimiza-

tion.

However, when we use M-CNN networks by mer-

ging the 3 different versions of the dataset into a 2-

channel architecture, we again explore the unique ex-

traction of characteristics from each version of the da-

taset, improving the learning of the model. Our appro-

ach has demonstrated that each version of the dataset

enriches the learning of the model, promoting a signi-

ﬁcant gain in accuracy.

Our results using the approach with multichan-

nel networks tacitly demonstrated that networks with

simpler architectures, such as AlexNet, obtained hig-

her accuracy to a network with denser architecture.

With only hyperparameter adjusting, our best single

channel result is better than 0.06% of the best result

obtained by Mohanty et al. (2016).

Considering the use of additional versions of the

dataset, grayscaled and segmented images, our best

result is better than 0.25% of the state-of-the-art va-

lue and it even outperforms our single channel met-

hod. The best combination output was with Version

1 of the dataset, that is colored and grayscaled images.

The overall results of M-CNNs were consistent with

any of the two different models, AlexNet and Goog-

leNet.

The graphics in Figures 3 up to 10 show that the

networks appear to stabilize after 30 epochs, but the

application of more epochs could increase the achie-

ved results.

Figure 3: Single channel networks accuracy on testing da-

taset.

Figure 4: Single channel networks losses on testing dataset.

5 CONCLUSIONS

In this work, we explored the potentialities of the

convolutional neural networks already evidenced by

the literature to identify plant diseases through sam-

ples from healthy and diseased plants. We explored

Plant Diseases Recognition from Digital Images using Multichannel Convolutional Neural Networks

455

Figure 5: Multichannel networks accuracy on testing data-

set for Version 1.

Figure 6: Multichannel networks losses on testing dataset

for Version 1.

Figure 7: Multichannel networks accuracy on testing data-

set for Version 2.

Figure 8: Multichannel networks losses on testing dataset

for Version 2.

Figure 9: Multichannel networks accuracy on testing data-

set for Version 3.

Figure 10: Multichannel networks losses on testing dataset

for Version 3.

primarily knowledge gaps highlighted by Mohanty

et al. (2016) optimizing and improving their results

and proposing an approach using convolutional neu-

ral networks with multichannels. The training of the

models was performed using an openly available da-

tabase PlantVillage, consisting of 54,306 images con-

taining 38 classes. We adopted the same strategy

as Mohanty et al. (2016) when performing the training

with three preprocessed versions of the PlantVillage

dataset deﬁned as color, grayscale and segmented.

In the ﬁrst step of our approach, we achieved sig-

niﬁcant advances as the accuracy of single channel

networks, optimizing the hyperparameters and adjus-

ting the dropout layers according to the dataset cha-

racteristics to minimize overﬁtting. It should be noted

that knowing the gains of transfer learning techniques,

we chose to train from scratch in order to demonstrate

the possibility of customization and gains in the lear-

ning process compared to a sample considered small

for a dataset.

Also, the additional inputs of the network provide

an even better accuracy, showing that M-CNNs were

able to enhance the general system, generating the

best overall result in this work and keeping the mean

scores regular and robust, independently of the

chosen model. The reference model achieved 0.9934

while our M-CNN obtained 0.9959. The dataset avai-

lable did not have images of plants in cultivated envi-

ronments, so the results of our approach only contem-

plate the tests performed extremely under preproces-

sed images and acquired in controlled environments.

Overall, we can conclude that a M-CNN model

trained from scratch is better than a single channel

model with transfer learning in two aspects: faster

convergence and reduced processing time. Further-

more, other image frequencies (e.g. grayscale) are

crucial to improve the general accuracy. Also, when

we train a single convolutional neural network from

scratch we achieve a model 10 times smaller than a

single channel with transfer learning. This reduction

allow us to build applications using real time plant di-

sease identiﬁcation in an open ﬁeld using mobile de-

vices.

In future works, the next step is to apply this ap-

proach to a dataset of images of healthy and disea-

sed plants obtained in growing environments. Thus,

it will be possible to adjust the M-CNN approach to

meet the requirements for identifying and classifying

plant diseases in culture from new real collected ima-

ges.

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

456

ACKNOWLEDGEMENTS

The authors also express their gratitude and ackno-

wledge the support of NVIDIA Corporation with the

donation of the Titan Xp GPU used for this research

and the ﬁnancial support received from the Federal In-

stitute of Education, Science and Technology of Mato

Grosso.

REFERENCES

Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z.,

Citro, C., Corrado, G. S., Davis, A., Dean, J., De-

vin, M., Ghemawat, S., Goodfellow, I., Harp, A., Ir-

ving, G., Isard, M., Jia, Y., Jozefowicz, R., Kaiser,

L., Kudlur, M., Levenberg, J., Man

e, D., Monga, R.,

Moore, S., Murray, D., Olah, C., Schuster, M., Shlens,

J., Steiner, B., Sutskever, I., Talwar, K., Tucker, P.,

Vanhoucke, V., Vasudevan, V., Vi

egas, F., Vinyals, O.,

Warden, P., Wattenberg, M., Wicke, M., Yu, Y., and

Zheng, X. (2015). TensorFlow: Large-scale machine

learning on heterogeneous systems. Software availa-

ble from tensorﬂow.org.

Al-hiary, H., Bani-ahmad, S., Reyalat, M., Braik, M., and

Alrahamneh, Z. (2011). Fast and accurate detection

and classiﬁcation of plant diseases. International

Journal of Computer Applications, 17(1):31 – 38.

Altieri, M. (2018). Agroecology: The Science Of Sustai-

nable Agriculture. CRC Press, Endereo, 2nd edition

edition.

Anderson, P. K., Cunningham, A. A., Patel, N. G., Morales,

F. J., Epstein, P. R., and Daszak, P. (2004). Emerging

infectious diseases of plants: pathogen pollution, cli-

mate change and agrotechnology drivers. Trends in

Ecology & Evolution, 19(10):535 – 544.

Arnal Barbedo, J. G. (2013). Digital image processing

techniques for detecting, quantifying and classifying

plant diseases. SpringerPlus, 2(1):660.

Baccouche, M., Mamalet, F., Wolf, C., Garcia, C., and Bas-

kurt, A. (2011). Sequential deep learning for human

action recognition. In International Workshop on Hu-

man Behavior Understanding, pages 29–39. Springer.

Chollet, F. et al. (2015). Keras.

https://github.com/fchollet/keras.

Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-

Fei, L. (2009). ImageNet: A Large-Scale Hierarchical

Image Database. In CVPR09.

Ferentinos, K. P. (2018). Deep learning models for plant

disease detection and diagnosis. Computers and Elec-

tronics in Agriculture, 145:311 – 318.

Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep

Learning. MIT Press. http://www.deeplearningbook.

org.

Hughes, D., Salath

e, M., et al. (2015). An open access re-

pository of images on plant health to enable the deve-

lopment of mobile disease diagnostics. arXiv preprint

arXiv:1511.08060.

Jabal, M. F. A., Hamid, S., Shuib, S., Ahmad, I., Jabal, M.

F. A., Hamid, S., Shuib, S., and Ahmad, I. (2013).

Leaf Features Extraction and Recognition Approa-

ches To Classify Plant. Journal of Computer Science,

9(10):1295–1304.

Ji, S., Xu, W., Yang, M., and Yu, K. (2013). 3d convolu-

tional neural networks for human action recognition.

IEEE transactions on pattern analysis and machine

intelligence, 35(1):221–231.

Karpathy, A., Toderici, G., Shetty, S., Leung, T., Sukthan-

kar, R., and Fei-Fei, L. (2014). Large-scale video

classiﬁcation with convolutional neural networks. In

CVPR.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012).

Imagenet classiﬁcation with deep convolutional neu-

ral networks. In Pereira, F., Burges, C. J. C., Bottou,

L., and Weinberger, K. Q., editors, Advances in Neu-

ral Information Processing Systems 25, pages 1097–

1105. Curran Associates, Inc.

LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep lear-

ning. Nature, 521(7553):436–444.

Lecun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998).

Gradient-based learning applied to document recogni-

tion. Proceedings of the IEEE, 86(11):2278–2324.

Lin, Z., Mu, S., Shi, A., Pang, C., and Sun, X. (2018). A

novel method of maize leaf disease image identiﬁca-

tion based on a multichannel convolutional neural net-

work. Transactions of the ASABE, 0(0):0.

Lu, Y., Yi, S., Zeng, N., Liu, Y., and Zhang, Y. (2017).

Identiﬁcation of rice diseases using deep convolutio-

nal neural networks. Neurocomputing, 267:378 – 384.

Mahlein, A.-K. (2016). Plant disease detection by ima-

ging sensors–parallels and speciﬁc demands for preci-

sion agriculture and plant phenotyping. Plant Disease,

100(2):241–251.

Miller, S. A., Beed, F. D., and Harmon, C. L. (2009). Plant

Disease Diagnostic Capabilities and Networks. An-

nual Review of Phytopathology, 47(1):15–38.

Mohanty, S., Hughes, D., and Salath, M. (2016). Using

deep learning for image-based plant disease detection.

Frontiers in Plant Science, 7(September).

Pan, S. J. and Yang, Q. (2010). A survey on transfer le-

arning. IEEE Transactions on Knowledge and Data

Engineering, 22(10):1345–1359.

Patil, S. B. and Bodhe, S. K. (2011). Leaf disease severity

measurement using image processing.

Pawara, P., Okafor, E., Surinta, O., Schomaker, L., and Wie-

ring, M. (2017). Comparing local descriptors and bags

of visual words to deep convolutional neural networks

for plant recognition. In Proceedings of the 6th Inter-

national Conference on Pattern Recognition Applica-

tions and Methods, ICPRAM 2017, Porto, Portugal,

February 24-26, 2017., pages 479–486.

Pydipati, R., Burks, T., and Lee, W. (2006). Identiﬁcation of

citrus disease using color texture features and discri-

minant analysis. Computers and Electronics in Agri-

culture, 52(1):49 – 59.

Rumpf, T., Mahlein, A.-K., Steiner, U., Oerke, E.-C.,

Dehne, H.-W., and Plmer, L. (2010). Early detection

and classiﬁcation of plant diseases with support vector

Plant Diseases Recognition from Digital Images using Multichannel Convolutional Neural Networks

457

machines based on hyperspectral reﬂectance. Compu-

ters and Electronics in Agriculture, 74(1):91 – 99.

Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S.,

Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bern-

stein, M. S., Berg, A. C., and Li, F. (2014). Image-

net large scale visual recognition challenge. CoRR,

abs/1409.0575.

Singh, V. and Misra, A. (2017). Detection of plant leaf di-

seases using image segmentation and soft computing

techniques. Information Processing in Agriculture,

4(1):41 – 49.

Sladojevic, S., Arsenovic, M., Anderla, A., Culibrk, D., and

Stefanovic, D. (2016). Deep neural networks based

recognition of plant diseases by leaf image classiﬁca-

tion. Computational Intelligence and Neuroscience,

2016.

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I.,

and Salakhutdinov, R. (2014). Dropout: A simple way

to prevent neural networks from overﬁtting. Journal

of Machine Learning Research, 15:1929–1958.

Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., An-

guelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.,

et al. (2014). Going deeper with convolutions.

Yosinski, J., Clune, J., Bengio, Y., and Lipson, H. (2014).

How transferable are features in deep neural net-

works? In Ghahramani, Z., Welling, M., Cortes, C.,

Lawrence, N. D., and Weinberger, K. Q., editors, Ad-

vances in Neural Information Processing Systems 27,

pages 3320–3328. Curran Associates, Inc.

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

458