International Journal of Circuit, Computing and Networking 2020; 1(2): 01-04

E-ISSN: 2707-5931
P-ISSN: 2707-5923
IJCCN 2020; 1(2): 01-04 An efficient approach for detecting grape leaf disease
Received: 01-05-2020
Accepted: 03-06-2020 detection
Pooja Amudala
GATE College, Tirupati, Pooja Amudala
Andhra Pradesh, India

Abstract
Grape diseases are the main causes leading to a serious reduction in grapes. Thus, it is urgent to
develop an automatic method of identification for diseases of the grape leaf. Deep learning techniques
have recently achieved impressive successes in various computer vision problems, which inspire us to
apply them to the identification of grape diseases. In this article, the convolution neural Networks
(CNN) The proposed CNN architecture. The combination of multiple CNNs enables the proposed
United Model to extract additional discriminatory features. The representative potential of United
Model has thus been enhanced. The United Model were evaluated on the Plant Village dataset and
compared to other state-of-the-art CNN models. Experimental results have shown that United Model
achieves the best performance in the various evaluation metrics. The United Model achieves an average
validation accuracy of 99.17 million quarter and a test accuracy of 98.57 per cent, which can act as a
decision support tool to help improve decision-making.

Keywords: CNN model, deep learning, grape diseases, united model, accuracy

1. Introduction
Plant disease has for quite some time been one in all the most significant dangers to food
security because of it drastically lessens the harvest yield and bargains its quality. Right and
exact diagnosing of maladies has been a significant test. Generally, the recognizable proof of
plant infections has depended on human comment by visual investigation. These days, it's
consolidated or subbed with fluctuated advances like immunoassays (e.g., chemical
connected immunosorbent measure, ELISA) and PCR or RNA-seq to watch pathogen-
explicit antigens or oligonucleotides, severally [1, 2].
Additionally, late classify advances and emotional value decreases inside the field of
computerized picture securing have permitted the presentation of a variety of picture based
diagnosing ways at a reasonable level [3]. In any case, in light of the fact that the no
inheritable picture encases dense data that is exceptionally irksome for the pc to strategy, it
needs a preprocessing venture to remove a positive element (e.g., shading and shape) that is
physically predefined by experts [4, 5]. In such things, profound learning is typically utilized
because of it allows the pc to self-ruling become familiar with the chief suitable element
while not human mediation. AN underlying arrangement to utilize profound learning for
picture-based ailment diagnosing was supposed in 2016, any place the prepared model had
the option to order fourteen harvests and twenty-six infections with a precision of ninety-
nine.35% against optical pictures [6].
From that point forward, consecutive ages of profound learning-based disorder diagnosing in
differed crops are supposed [7 – 13]. Among fluctuated organize designs utilized in profound
learning, convolutional neural systems (CNN) square measure wide utilized in picture
acknowledgment. The essential CNNs, the notoriety [14] and LeNet [15], were presented inside
the Eighties, however the investigation of neural systems initially began inside the Nineteen
Forties. CNN's are utilized for plant picture examination since the primary days of their
advancement. As a result of the expedient advancement of equipment and furthermore the
improvement of learning ways enormous scope profound CNNs got trainable inside the 2010
a genuine turning reason for the CNNs was the presentation of Alex Net, that significantly
beat the picture arrangement precision of antiquated AI approaches in ImageNet gigantic
Scale Visual Recognition Challenge (LSVRC) 2012. CNN's includes convolutional layers,
Corresponding Author:
Pooja Amudala that square measure sets of picture channels tangled to pictures or highlight maps, along the
GATE College, Tirupati, edge of option (e.g., pooling) layers. In picture grouping, highlight maps square measure
Andhra Pradesh, India removed through convolution and elective procedure layers

~1~
International Journal of Circuit, Computing and Networking http://www.computersciencejournals.com/ijccn

redundantly and furthermore the system in the long run to each pixel and then selecting a single value (eg.
yields a name showing a measurable classification. Given an maximum) from with the mask.
instructing dataset, CNN, as opposed to old Machine
Learning methods that utilization overhand alternatives, Experimental and technical design
improves the loads and channel boundaries inside the To reveal the characteristics of visual image approaches for
concealed layers to get choices suitable to determine the CNN for disease detection, we tend to adopt Nemours way
characterization downside. In principle, the boundaries on a trained CNN model employing a plant disease dataset.
inside the system square measure streamlined by back- We tend to competed four class of visualization image ways
proliferation and slope plunge approaches [14, 15] to weaken 1. Hidden layer output visual image
the classification error. 2. Feature visual image 3) linguistic wordbook and 4)
Attention maps.
2. Related works
Machine learning may be a machine manner of detection Dataset and network for disease diagnosis training
patterns in an exceedingly given dataset so as to create This dataset includes healthy or pathologic leaf pictures
inferences in another, similar dataset. A classical textbook classified into thirty eight labels (54,306 images, twenty six
example is that the machine recognition of handwriting like diseases, fourteen crop species, pictures were split into
communicating addresses on envelopes [8]. In recent years, coaching, validation, and take a look at datasets with a
the generic visual perception has created tremendous magnitude relation of 6:2:2. victimization such pictures, we
advances and is currently approaching human accuracy. tend to ready a CNN supported InceptionV3 that receives a
Within the paper, author Mrunalini represents the technique three-channel input image of 224 x 224 resolution and
to classify and establish the various diseases through which returns a 38-dimensional vector. We tend to selected this
plants area unit affected. Within the Indian Economy, a specification as a result of it's comprised of repetition
Machine learning-based recognition system can sway be convolution blocks while not advanced layers like residual
terribly helpful as it saves efforts, money, and time too [9]. connections which will build the interpretation of the
The approach given during this for feature set extraction is intermediate layers tough.
that the color co-occurrence method. For automatic
detection of diseases in leaves, the neural networks area unit Visualization: hidden layer output visualization
used. The approach projected will significantly support a We originally utilized one in everything about principal
correct detection of the leaf, and appears to be a very credulous manners by which to find out the educated
important approach, just in case of steam, and root diseases, choices and to extricate the shrouded layer yield (i.e.,
swing fewer efforts in computation. In paper they transitional yield); we tend to pass an image to the CNN and
incorporated all the hybrid options of a leaf color, texture stopped the computation at the layer of intrigue. Since a
form (geometric feature) by the individual methodology. component extraction layer passes exclusively the positive
Plant Village: a tool for crop health; an internet platform qualities to the proceeding with layer because of our system
dedicated to crop health and crop diseases, referred to as applies the rectified direct measure (ReLU) activation
Plant Village [11]. The content has been written by plant operate, only visualize the intermediate output will be the
pathology consultants; reflective data is sourced from the middle of the road yields will give an unpleasant execution
scientific literature. of "What a piece of the picture was essential for the
deduction. As per the implementation connected work [6]
3. Proposed system explicitly focused on the yield of the essential convolutional
To reduce this loss of the crops production we present one layer, though we tend to utilized indistinguishable method
android app which recognize and distinguish the indications for each layer output.
of disease on plant leaf. Our application chip away at such
plants which are infected by any disease that is fungi, 4. Results and discussions
viruses, and classify plant disease by utilizing Deep learning
strategies. It identifies the real sort of disease and gives its
preventive measures and related recovery notations are
displayed by using CNN. And finally, we get information
regarding that disease its symptoms, its preventive
mechanism and recovery suggestions in a more economical
way.

Algorithm
In our previous work, we developed a CNN model, which
achieved accuracy greater than 93% for 15 different plant
types. This study will investigate the model in more detail.
Once images are read, 256x256 pixel random parts of the
images are extracted and noise, distortion, flip, or rotation Fig 1: High resolution Input Image
transforms are applied. By controlling stride lengths
(spacing interval for placement of the filters/masks), This is the input images of grape leaf which consists of
dimensions of masks, multiple convolutions, and pooling diseases. We are using these images for to detect whether
steps are applied. Pooling involves the application of a mask they are having any disease or not.

~2~
International Journal of Circuit, Computing and Networking http://www.computersciencejournals.com/ijccn

Fig 2: The Illustration workflow of the methodology

Total work flow of the model from input image to label the process we are getting important features through
extraction. Here we are giving input as the images and after feature extraction process.

Fig 3: Validation accuracy validation loss

This image discusses about the various accuracy of the different classes in the model.

Fig 4: Average Accuracy of the Different Architectures

~3~
International Journal of Circuit, Computing and Networking http://www.computersciencejournals.com/ijccn

Here we are fallowing different architectures accuracy. By 9. Athanikar G, Badar Potato P. Leaf diseases detection
comparing all the algorithms, we are getting the kinds of and classification system Int. J Comp Sci Mob Comput.
results about the grape leafs. In Fig 4 comparison curve 2016; 5(2):76-88.
representing the accuracy and loss of the training and 10. Xie D, Zhang L, Bai L. Deep learning in visual
validation evaluation results. computing and signal processing Appl Comput Intell
Soft Comput, 2017, 1-13.
5. Conclusion 11. Ouyang W, Wang X. Joint deep learning for pedestrian
In this work, we have developed an effective solution of detection Proc. IEEE international conference on
rapid recognition of grape disease based on CNNs. We have computer vision. Sydney, Australia, 2013, 2056-2063.
proposed the United Model, which is a unified CNN 12. Sun Y, Wang X, Tang X. Hybrid deep learning for face
architecture based on InceptionV3 which ResNet50, and can verification IEEE Trans Pattern Anal. Mach Intell.
be used to classify grape images into four classes, including 2016; 38(10):1997-2009.
three different symptom images, i.e. black rot, esca, 13. Zhang L, Yang F, Zhang Y, Zhu Y. Road crack
Arabidopsis leaf mark, and safe images. United Model is detection using deep convolutional neural network
going to take the advantage of combining the width of Proc. IEEE international conference on image
InceptionV3 and the depth of ResNet50 is that the more processing. Arizona, USA, 2016, 3708-3712.
representative features can be trained. The representational 14. Zhou Z, Shin J, Zhang L, Gurudu S, Liang J. Fine-
capacity of United Model is improved by a high-level fusion Tuning convolutional neural networks for biomedical
feature, which allows it to achieve the best performance in image analysis: actively and incrementally Proc. IEEE
the identification of grape diseases. conference on computer vision and pattern recognition.
Experiment results show that the model can outperform the Honolulu, Hawaii, 2017, 7340-7349.
state-of-the-art simple CNNs, including VGG16, 15. Mohanty S, Hughes D, Salathé M. Using deep learning
InceptionV3, DenseNet121, and ResNet50. The proposed for image-based plant disease detection Front Plant Sci.
United Model achieves an average validation accuracy of 2016; 7:1419.
99.17% and a test accuracy of 98.57% and can therefore act 16. Lottes P, Behley J, Milioto A, Stachniss C. Cyrill Fully
as a decision support tool to help farmers identify grape convolutional networks with sequential information for
diseases. We also include a realistic analysis to resolve the robust crop and weed detection in precision farming
inadequacy and inconsistency of the data collection. IEEE Robot Autom Lett. 2018; 3(4):2870-2877.
Methods of data increase strategies, early stop mechanism 17. Wang X, Cai C. Weed seeds classification based on
and drop-out are used to boost the model's ability to PCANet deep learning baseline Proc IEEE Asia-Pacific
generalize and the chance of over-fitting. In addition, we signal and information processing association annual
have introduced an efficient multi-network integration summit and conference. Hong Kong, China, 2015, 408-
approach that can be used to incorporate further state-of-the- 415.
art CNN. 18. Cheng X, Zhang Y, Chen Y, Wu Y, Yue Y. Pest
identification via deep residual learning in complex
6. References background Comput Electron Agric. 2017; 141:351-
1. Kole DK, Ghosh A, Mitra S. Detection of downy 356.
mildew disease present in the grape leaves based on 19. Rahnemoonfar M, Sheppard C. Deep count: fruit
fuzzy set theory Advanced computing, networking and counting based on deep simulated learning Sensors.
informatics, Springer, Berlin (Germany), 2014, 377- 2017; 17(4):905 Google Scholar.
384. 20. Kussul N, Lavreniuk M, Skakun S, Shelestov A. Deep
2. Martinez A. Georgia plant disease loss estimates. Link: learning classification of land cover and crop types
http://www.caes.uga.edu/Publications/displayHTML.cf using remote sensing data IEEE Geosci Remote S.
m?pk_id=7762, 2015. 2017; 14(5):778-782.
3. Srdjan S, Marko A, Andras A, Dubravko C, Darko S.
Deep neural networks based recognition of plant
diseases by leaf image classification Comput Intell
Neurosci, 2016, 1-11.
4. Wang G, Sun Y, Wang J. Automatic image-based plant
disease severity estimation using deep learning comput
Intell Neurosci, 2017, 1-8.
5. Xu G, Yang M, Wu Q. Sparse subspace clustering with
low-rank transformation Neural Comput Appl, 2017, 1-
14 CrossRefView Record in ScopusGoogle Scholar.
6. Singh V, Misra A. Detection of plant leaf diseases
using image segmentation and soft computing
techniques Inf Process Agric. 2017; 4(1):41-49.
7. Ma J, Du K, Zheng F, Zhang L, Gong Z, Sun Z. A
recognition method for cucumber diseases using leaf
symptom images based on deep convolutional neural
network Comput Electron Agric. 2018; 154:18-24.
8. Xu X, Liu Z, Wu Q. A novel K-nearest neighbor
classification algorithm based on maximum entropy Int.
J Adv Comput Technol. 2013; 5(5):966-973.

~4~