On detecting plant diseases with image p

Through this particular survey paper and its findings, we have tried to show various types of image processing techniques such as feature extraction, segmentation, classification techniq

On Detecting Plant Diseases with Image

Processing: A Survey

Epsita Medhi

Department of Information Technology

Gauhati University

Assam, India

epsitamedhi12@gmail.com

Nabamita Deb Department of Information Technology

Gauhati University Assam, India deb.nabamita@gmail.com

ABSTRACT

With the modernization of the agricultural industry, applications

of image processing have a considerable impact on this industry

Image processing applications have been used in the early and fast

detection of crop and plant diseases The early detection of

diseases enables the people concern to treat the infected crops or

plants which results in ensuring good quality and yield of the

crops and plants being raised by the farmers This paper describes

the recent studies, techniques, and contributions made by different

researchers, to detect plant diseases affecting various plant parts

Current image processing trends and techniques like K-Means,

Histogram Equalization, Otsu method, Adam Optimizer, etc have

been put forward in this survey paper to serve the purpose of

identifying and detecting diseases and their pathogens The study

also discusses the various diseases and the corresponding image

processing applications used by the researchers to identify the

various diseases The goal of this paper is to identify the current

trends of the applications of image processing methods to

recognize various plant diseases Techniques like GLCM,

GABOR filter, Local Binary Pattern have been used by

researchers for the extraction of the features from the images

K-Means, Fuzzy logic, Otsu, etc have been used for segmentation

and clustering Some of the survey papers include SIFT,

Euclidean, and Autoencoders as classifiers

Keywords

Gabor filters, Image classification, Image enhancement, Filtering

techniques, Segmentation, Image processing, Feature selection

1 INTRODUCTION

India being an agricultural country, its highest economic income

is based on the agricultural yields that are produced [1] With its

proper irrigation facilities, fertile soil, hard labor, India produces

large amounts of crops and plantations The agricultural outputs

get hampered because of various reasons such as floods, drought,

etc but the most important of them all is due to the attack of the

pathogens and disease-causing microorganisms To detect the

diseases at their onset would be much beneficial for the protection

of the crops

Image processing techniques help in solving various

problems by applying their different algorithms It can also

remove unwanted noises and signals A combination of user's

data, knowledge, and processed data can help in solving various

problems through image processing

The automation techniques and tools can be utilized by

the farmers for improving the productivity, grade, and yield to

make use of smart farming From time to time investigation of the

agricultural field along with the early detection of the diseased

crops help in the growth of healthy crops along with the profit in

the economy Various techniques have been formulated to support

this desired result [2]

Farmers, agriculturists, horticulturists, etc usually

detect the diseases by looking at them And at many times

agricultural experts are called upon to validate the same By doing

this, there comes the factor of heavy costs, and also the constraint

of time

Through this particular survey paper and its findings,

we have tried to show various types of image processing techniques such as feature extraction, segmentation, classification techniques, to name a few to detect plant diseases and show their classifications

2 BASIC STEPS IN IMAGE PROCESSING FOR DETECTING DISEASES CAUSED IN PLANTS

Image processing helps in the improvement of accuracy and the consistency of some agricultural practices, thereby helping the farmers to produce good quality yields The automated techniques help in taking quality measures which is sometimes more beneficial than visual decision making [3] The steps involved are discussed as follows At first, images are acquired with the help

of a digital camera which can also be a mobile camera Acquired images are divided into two datasets that are healthy and diseased types Healthy plant images are kept so that we can easily compare them with the diseased ones Preprocessing can be performed on the diseased images by applying some techniques like geometric transformation, filtering, brightness correction, etc Region of interest can be calculated with the help of segmentation techniques like thresholding, histogram method, etc

Extraction of the features from the selected region of the images can be achieved using independent component analysis, PCA, LBP techniques Symptoms of the diseases, categories of the diseases can be measured or categorized with the help of PCA, filter method, wrapper, and so on Classification of the categorized diseases can be completed with some techniques such as linear regression, KNN, SVM, etc And in the final process images can be identified as well as the diseases can be detected with the use of some techniques such as Fast R- CNN, RCNN, Single Shot Detector, etc The Flow Chart of the above-discussed processes is given below

Figure 1: The Steps involved in Detecting and Classifying plant diseases 3 LITERATURE REVIEW There has been extensive work by various researchers which has been included in this survey paper to show the use of the different techniques for classifying and detecting diseases in various parts of the plant Different types of pre-processing techniques, feature extractions, segmentation techniques, classification techniques, clustering techniques, etc have been applied by different Researchers Gavhale et al.[4]in their proposed method used SF-CES for enhancing the image and conversion has been done using the color space method Image segmentation by applying K-Means technique and the feature extraction by using GLCM method have been performed SVM technique has also been implemented It was found that the SF-CES had better color enhancement For disease part extraction, Lab and YCbCr support K-Means clustering Prakash et al.[5]in their proposed method have analyzed and classified different techniques for detecting diseases in plant leaves A color conversion has been done For clustering and segmenting, the K-Means method of clustering has been used Extraction of features has been done using GLCM They have also used an SVM classifier Kernel Hilbert Space has been applied for reducing computational complexities Through their proposed method, they could easily distinguish the healthy and diseased parts of the plant Zhang et al.[6]proposed an intelligent fruit detection system method by improving the Multi-Task Cascaded Convolutional Neural Network (MTCNN) AuRo (Automated Robot) has been working in the real-time and with higher accuracy A procedure for a fusion of the images has been used for improving the performance of the detector The proposed detector worked without any error and had a good time-cost and accuracy Combining negative patches and samples from the dataset, a Fusion Algorithm has been generated by them The technique proposed works well with other convenient objects and methods They used PNet as the convolution layer Dandawate et al[7] helped farmers over internet by developing a Decision Support System (DSS) They have used a Support Vector Machine (SVM) RGB image formats are converted to HSV color space during pre-processing A multi-thresholding technique has been used to extract the ROI OTSU technique and SIFT technique were used for classification and recognition The segmentation technique was based on color and cluster Automatic recognition is done using Scale Variant Feature Waghmare et al.[8] in their proposed method have developed a technique for the detection of leaf diseases They carried out their work by texture analysis of leaf and recognizing the pattern They have performed segmentation after removing the background Different autonomous diseases will have different textures and then the pattern is classified by multiclass type SVM Textures were analyzed using Local Binary Pattern Through their proposed method, identification of diseases can be done in terms of shape, texture, and color The texture is analyzed using Opposite Color Local Binary Pattern Classification performed using SVM Rastogi et al.[9]in their paper have proposed a leaf disease identification technique using a computational scenario Two-phase implementation has been done in their proposed method The first phase includes preprocessing, feature extraction Feature extraction by GLCM matrix and the Artificial Neural Network was applied in their system In the second phase, classification has been done using K-Means classifier and feature extraction done in the Region of Interest Grading of the disease has been done K-Means clustering and Euclidean technique have also been applied Their system provides immense help to the farmer for automatically detecting the diseases in the leaves of the plant Luna et al[10]developed an effective solution for automatic disease detection in plants of tomato with a motor-controlled capturing box of images GUI, DCNN were developed together Deep Learning Algorithm with an autoencoder was developed for the classification of images and in learning future tasks The actual results of the recognition of tomato plant detection were found to be successful if the diseases matched the diagnosis of the agriculturists Akhtar et al.[11]in their paper compared various machine learning performances for identifying and classifying leaf diseases OTSU algorithm has been used for selecting the threshold value According to their experiment, the decision tree is the best classifier and the best performance is given by the Discrete Wavelet Transform (DWT) They found that when they combined the DCT and Decision Tree, their system yielded much better accuracy And when DCT, Decision Tree, and SVM were combined their system gave the best performance Their proposed technique was better than the other classification techniques Jhuria et al.[12]in their paper, proposed a detection and classification process for detecting diseases in the plant leaves They considered ANN and Backpropagation method For clustering, the K-Means technique was used Fuzzy logic has been used for the automatic grading of fruit By using their software tool, farmers would be much benefitted and help in checking the agricultural yield from time to time Kulkarni et al.[13]proposed a method to detect early diseases in pomegranate plants CIELAB has been used for segmenting the uniform color scale Here features extraction has been done using GABOR filter and classification has been done using ANN classifier, which gave a better result Khampria et al [14]designed a hybrid approach by combining the Deep Convolutional Neural Networks and Autoencoders to detect diseases in crop leaves Application of Adam Optimizer for accuracy, ReLu for activation function, and Back Propagation for measuring weights can be seen in their experiment The accuracy in the number of epochs changes with different convolutional filter sizes, which resulted in better performance They had a smaller dataset and because of the lack of good GPUs, the training became very large PREPROCESSING SEGMENTATION OF ROI MOBILE CAMERA/ DIGITAL CAMERA IMAGE ACQUISITION HEALTHY PLANTS (LEAF/ STEM) DISEASED PLANTS (LEAF/ STEM) GEOMETRIC TRANSFORMATION, FILTERING, BRIGHTNESS

CORRECTION, etc

THRESHOLDING, EDGE BASED

METHOD, HISTOGRAM METHOD, etc

INDEPENDENT COMPONENT

ANALYSIS, PCA, LBP, GLCM, etc

FEATURE EXTRACTION

PCA, FILTER METHOD,

WRAPPER, EMBEDDED, etc

SYMPTOMS/ DISEASE CATEGORY

LINEAR REGRESSION, KNN,

DECISION TREE, SVM, etc

CLASSIFICATION

DETECTION OF DISEASES FAST R-CNN, HOG, RCNN, SINGLE

SHOT DETECTOR, etc

Le et al [15]used features that were extracted by combining

operators of Local Binary Patterns along with the extracted

features from plant leaves to distinguish between broadleaf plants

Plant contour masks were used for improving the rate of

discrimination between the plants comprising of broadleaves For

filtering the noise, operators of morphological characters were

used Classification is done using SVM The system proposed can

distinguish similar types of crops and can also detect weeds

Petrellis [16]proposed a similar technique to detect human and

plant diseases OCTAVE method has been used in the system

Fuzzy logic has been applied for performing segmentation and

classification Gaussian Low Pass filter has been employed for

MRI scan Classification uses the ANN classifier Supervised

clustering in the system yields a better result than unsupervised

learning

McDermott et al [17]has provided an overview of the methods

that were employed for computational predictions of the effectors

SIEVE server has been used as a web-based tool The

classification has been done using SVM On comparing Sequence

Order Independence (SOI) with Sequence Order Dependence

(SOD), it was found that SOI is better ROC AUC classification

was found to be good

Amara et al.[18]in their paper, proposed a deep learning classifier

that classifies the predictive performance of unseen diseases of

banana leaves LeNet architecture under CNN has been used for

image classification OTSU method was used to focus the ROI

against the background To identify infected areas, the K-Means

clustering technique has been used The model works as a

decision support system to identify plant disease

Sperschneider et al.[19]predicted an effector in fungi for

conserving features of motifs in the sequence of N-terminal

EffectorP has been introduced for learning the machine

application to predict effectors of fungi The EffectorP has an

accuracy of over 80% which helps in the prediction of fungal

effector based on Secretomes To predict candidates of the

effector with a high-priority, EffectorP gets combined within

planta expression for more power Feature selecting strategies

applied like Exhaustive, Hill-climbing Greedy searches which are

a selection of features correlated provided by WEKA EffectorP

predicts effectors of species-specific and core types The

biological mechanism can be understood with its help

Dheeb et al.[20] in their paper, proposedmethods to study, design,

implement and evaluate plant leaf disease for its automatic

detection and classification For that, they have divided their

system into phases Clustering has been done by the K-Means

algorithm Texture features were calculated using the Color

Co-occurrence Method Neural Network Classifier that was

developed performed well and successfully tested the system It

was also observed that misclassification mainly occurred in the

four classes that they have taken into consideration

Sladojevic et al.[21]developed a new method for the recognition

of plant leaf disease with DCNN Different color models such as

YCbCr, HIS, CIELAB have been used for the study Their model

studies the trained network to distinguish features from one

another Translation and rotation have been done using Affine

Transformation For measuring the weights CaffeNet was

applied CNN and ReLus were applied in their system for

convolutional and fully connected layers When compared with

other models it was found that their model yielded better results

Ramesh et al.[22]in their paper, proposed method in which

classification cum recognition of diseases in crop leaves has been

done Their model removes background noise and segmentation

for clustering the diseased portion Clustering using K-Means

done to the hue part of HSV Texture features were extracted

through GLCM Weights of the nodes were updated by the

combination of JOA and DNN iteratively in the hidden layers

DNN_JOA classifier was found better than to rest of the

classifiers and it achieved the highest accuracies

Li et al.[23]in their paper, proposed video detection architecture

for detecting real-time crop diseases along with pests They have

proposed custom CNN, Fast-RCNN, R-CNN, Faster-RCNN for

detecting objects They have also proposed a custom DCNN backbone for Faster-RCNN In their model, they have used the still-images detection metrics related to True Positive and Negative, etc for calculating the lesion spots For converging images and videos Stochastic Gradient Descent (SGD) becomes handy DCNN backbone develops a good result in comparison to VGG, ResNet-50, ResNet-101 Detection from videos caused some problems like Video Defocus with Part Occlusion and Motion Blur Lesion spot shape was found to be always irregular Mohanty et al.[24]developed a model based on a smartphone that helps in assisting the disease diagnosis They used Stochastic Gradient Descent as parameters for performance Alex Net, Google Net were newly designed by them and found that Google Net performed better than Alex Net and segmentation showed variation in performance It was found that the model performed better for the color version of the dataset

Dey et al.[25]in their paper, proposed a technique in detecting pests on leaves of the various plants K-Means technique has been used in their model GLCM, GLRLM techniques have been applied to extract the features Their proposed method uses some pre-processing tasks such as Noise Removal, Image Contrast Improvement, etc From their experiment, it was found that the SVM classifier achieved the highest accuracy

Keh [26]through his proposed model, investigated the classification of pathology problems with single leaf images They developed a new Efficient Net model Accuracy measured using Adam Optimizer The losses in models were combined to compensate for the shortcomings It was found that their new model Efficient Net performed better than the other models It was also found that if Efficient Net and Noisy Student Training were combined then the system gave a much better result Chohan et al.[27]in their model, proposed CNN architecture for detecting diseases in plants Augmentation increases the dataset size Feature extraction has been done on horizontal edges, vertical edges, RGB images From their experimental model, it was found that CNN predicted the diseases correctly on plant leaf images

Minaee et al.[28]provided comprehensive literature review for segmentation, including convolutional networks They covered literature for the segmentation of images and discussed Segmentation methods based on deep learning In the case of survey papers, as mentioned by them, various techniques such as AlexNet, VGGNet, ResNet, GoogleNet, MobileNet, and DenseNet have been applied LSTM (Long Short Term Memory) has been used by them They provided different aspects for the future which would be challenging for segmentation, based on deep learning

Lin et al.[29]proposed a new model for segmentation of semantic criteria concerning CNN for detecting powdery mildew disease,

at the pixel points, in the leaves of the cucumber Image augmentation technique and custom loss function along with its layer of batch normalization have been added with the convolutional layers OTSU method was used for masking ADAM method for optimization and Glorot initialization method for initializing the weights were used by them Their proposed method for clustering was better than K-Means They found that their proposed method in terms of Precision was not good, but in terms of Recall, it was found to be good Their method could be easily implemented and was rather cheap They also found that along with the diseased region, some non-diseased regions have been identified as diseased regions

Iqbal et al.[30]in this survey paper showed the different classification and detection of disease techniques in the leaves of

a citrus plant They surveyed different terms and techniques that have been used by different researchers for carrying out their research work such as Otsu Thresholding, Edge Detection, Compression Based Methods, etc It also discussed the importance of extraction of features along with deep learning-based methods They proposed efficiency improvement in terms

of color space, YCbCr, images, along with the K-Means clustering algorithm Texture calculated using GLCM For

classification, the backpropagation technique was considered

The papers that were surveyed by them showed that none of the

researchers used saliency-based techniques which works well

with the detection of diseases FS and DL methods achieved high

accuracy and perfect classification time

Raufa et al.[31]through their method included four main

processes such as dataset enhancement, segmentation of the

lesion spot, feature extraction from the region infected, and

classification of the infected region Filtering of the images done

through Top-hat processes and for improving the contrast,

Gaussian function has been applied Images that were enhanced

are then mapped to the saliency graph Geometric features, color,

textures were extracted using Entropy, PCA, Skewers The

top-hat process was applied to the original images and Gaussian

Function was applied for better contrast Finally, the classification

being performed with each instance of images in correspondence

to classes of each disease Identification of diseases with naked

eyes becomes a negative factor because of weather and light

Thomas et al [32]in their paper have checked Phenotyping as a

narrow observation for cultivars development This study

provides a better system of hyperspectral phenotyping, that

combines measurements of canopy scale with the environment of

higher spatial resolutions Symptoms of powdery mildew were

detected by combining Simplex Volume Maximization and SVM

Savitzky-Golay filter has been used for smoothening of the

normalized images Under artificial light, hyperspectral images

acquired superior quality than that of images acquired under

natural light

Pugoy et al.[33]proposed disease detection system in the leaves

of rice crop with images having color analyzer A region

including the outlier obtained with histogram intersection

between healthy and test images of the leaves of rice For

clustering, K-Means was applied and from that, the diseases can

be determined Their system works on Bhatia's algorithm where

pixels are selected and clustered using Euclidean distance Their

proposed system can inspect the rice leaf diseases by comparing

and matching colors

Sujatha et al.[34]have classified citrus leaf disease using both the

methods of deep learning and machine learning to check which

methods will be most feasible They applied cross-validation with

10-fold for classification problems and each of the fold consists

of the same ratio as that of each class of target Image prediction

is done when input is given Methods of Random Sampling and

Cross-Validation have been done to train and test the system

They used Squeeze Net (SN) for the embedding of images For

image classification SVM was being considered by them For

optimizing an objective function, they used Stochastic Gradient Descent They have considered 10 Random Forest trees that can replicate and control growth Inception V3 has been used to increase the accuracy and reducing overfitting for fully connected layers VGG-16 was also studied for padding, max pooling

VG-19 has been used for building small-sized convolution, for deep neural networks On comparing the works of DL with ML it was found that the techniques of Deep Learning worked better than that of Machine Learning

Xiao et al [35] developed a technique for recognizing images of disease in strawberry plants They have considered CNN techniques and a ResNet50 model to detect leaf blight Average training accuracy and loss rates were calculated for the original and feature image datasets They have worked with VGG-16, GoogleNet, ResNet-50 and have compared their accuracy with each one With their method, the diseases of the strawberry plant were properly detected

TABLE 1: SUMMARY OF THE LITERATURE REVIEW PAPERS:

SL

N

O

REFERE-NCE

YEA-R

SUMM-ARY

DATA-SET

METHODOL-OGY

RESULTS/

VISUALIZA-TION/

ACCURACY

ADVANTAG-ES/ DISADVANT-AGES

1 Gavhale [4] 2014 Development

of a model to detect diseases in leaf through image analysis along with classification techniques

Total samples:

250 images

Classes:

Canker, Anthracnos

e

SF-CES to enhance

the image and converting color

space, K-Means for

segmentation,

extracting features, along with it, classification have been done

The contrast of the images is constant when the value is 0

between 0 and 1, homogeneity between 0 and 1, a correlation between 0 and -1 are set

Different rates of acceptances have been calculated

SF-CESs color enhancement was found to be better The genuine Acceptance Rate was better than the others

2 Prakash [5] 2017 Implementati

on of a

Total samples:

To determine chromaticity and the

Features are extracted

Proposed method for citrus leaf detection

technique for analyzing images and classifying

detect diseases in the leaves

Leaves of citrus: 60 collected images

Classes:

Diseased 35 images, Healthy 25 images

luminosity layers, images in RGB format converted to

L*a*b K-Means for

clustering, GLCM

for extracting the

features, SVM for

classifying the images For reducing computational

complexities, Kernel Hilbert Space (RKHS) has been

used

homogeneity, energy, correlation, etc

Classifier’s performance

implementing actual cum predicted values

worked well

3 Zhang [6] 2017 Design of an

intelligent fruit detection system (InFD) by improving the Multi-task Cascaded Neural Network (MTCNN) to Fruit-MTCNN (F-MTCNN)

Total samples:

1800 images

Classes:

Apple, Strawberry, Oranges

Artificial image samples were generated using the

Fusion Algorithm

by adding negative patches from samples Feature extraction done using

PNet architecture of

3 CNN layers, false candidates are

removed by RNet and ONet The loss

function is calculated

by dividing it into two parts: Fruit Classification and Bounding Box Regression

AuRo worked in real-time with higher accuracy, the augmentation method was improved, 7 groups of negative

generated Using Fusion

Augmentation, the true positive rate can

be improved

The system worked properly in terms of accuracy and time cost It can detect any fruit

4 Dandawate[7] 2015 Proposed a

tool for detection of diseases in soybean plants The tool helped farmers over the internet

by developing a Decision Support System (DSS)

Total samples:

120 images

As the images are sent to the central system, the normal and the diseased leaves have been classified on features that were extracted

The images in RGB format converted to

HSV, segmentation using Otsu method,

recognition of plant

done through Scale Invariant Feature Transform

Classification and extraction of features

done through SVM

transformed via Scale Invariant Feature Transform SVM classifier gives better accuracy

Classification accuracy also improves It is found that the sample leaf generated from the graph is the exact match of the referenced leaf The accuracy achieved is

98.42%

SIFT algorithm helps

in matching species of the plant with leaves

5 Waghmare [8] 2016 Proposed a

model for identifying plant diseases

by recognizing the texture and pattern of the leaves of

a plant

Total samples:

450 images

Classes:

Grape diseases:

Black Rot, Downy Mildew

Background and unwanted images are removed using segmentation, textures are analyzed with the help of

Local Binary Pattern with Opposite Color,

classification is done

using SVM

Accuracy has been achieved with SVM multiclass classifier

On increasing the training and testing ratios, accuracy also increases The accuracy achieved is

96.66%

Automated Decision Support System proved to be advantageous

6 Rastogi [9] 2015 Proposal for

the development

computationa

l method that would help in the disease identification

Classes:

Leaf Scorch, Leaf Spot found in Maple and Hydrangea

Feature extraction

using GLCM matrix,

classification using

ANN, Segmentation

using K-Means

method Toolbox of Fuzzy Logic used comprising of

Different types of the cluster are generated

on loading the system such as Grayscale cluster, Cluster of the leaf that are segmented,

Clustering portion of diseases, and

The proposed method yields a good result

in leaves calculation of the

infection’s percentage

background of an input image

7 de Luna [10] 2018 Development

automatic solution for disease detection by a motor-controlled capturing device of image

Total samples:

4923 images

Classes:

Diamante Max, a breed of Tomato

The network is trained with 5 different layers such

as max-pooling, dropout,

convolutional layers

Data annotation was applied to avoid overfitting AlexNet architecture was modified Anomaly detection and disease recognition done for epochs of 50 with FRCNN

On adjusting SGD using 50 epochs, the confidence score acquired a good result Recognition accuracy was found

to be better For 36 samples, the accuracy was good The accuracy achieved is

91.6%

The proposed tool worked with good accuracy

8 Akhtar [11] 2013 Performance

comparison

of various techniques of Machine Learning to classify, identify diseases in leaves

Total samples:

40 images

Classes:

Black Spots, Anthracnos

e

Grayscale thresholding for segmentation, tiny dots removed using

open and fill operators, threshold

value selected using

Otsu method

Feature extraction using Statistical features, Discrete Cosine Transform, and Wavelet Transform

Classification is done using a Decision Tree

DCT and Decision Tree on combining gives good accuracy

DCT, DWT, SVM on combination gives the best accuracy The accuracy achieved is

94.45%

techniques worked with good results

9 Jhuria [12] 2013 Development

of a tool to monitor the diseases in fruits

Total samples: 92 Classes:

Black Rot and Powdery Mildew of Grapes and Apple Scab and Rot of Apples

Images are classified based on color, texture, and morphology, texture feature using

Daubechies 2-D wavelet packet decomposition,

backpropagation method used for recurrent networks,

K-Means clustering

has been used

Square Error Concept was used for training, Fuzzy logic for automatic

grading

For a different percentage of diseases, different grading has been found out Grading has been done by considering the ratio between the area of diseases and the entire area of the leaf

technique will give good results and benefits to the farmers

10 Kulkarni [13] 2012 Early

detection of diseases in pomegranate plants

Total

samples-140

Classes:

Alterneria-8 images, BBD-26 images, Anthracnos

e – 89 images

Segmentation using

CIELAB, Filtering using Gabor filter,

Classification using

ANN classifier

Optimum network efficiency and optimum termination error rate vs neural network efficiency have been achieved

achieved is 91%

The proposed system yields a better result

11 Khamparia

[14]

2019 Design of a

hybrid approach for detection of diseases in leaves by

Total samples

-900

Classes:

Potato:

High dimensional feature vectors at multiple levels

Autoencoders for learning encodings

Adam optimizer for

Different sizes of filters and epochs have variations in frequency levels with different accuracies

The proposed system yielded better results but GPU was small and so was the database

combining DCNN and Autoencoder

s

Early blight, Late Blight, Tomato:

Leaf mound, Yellow Leaf Curl, Maize:

Rust disease

increasing accuracy

ReLu has been used

as the activation function

Backpropagation algorithm for weights training

12 Le [15] 2020 Identification

of similar types of leaves specially broadleaves

improve the discriminatio

n rate using Plant Contour Masks

Total samples –

15000

Classes:

Canola –

7500 images and Radish – 7500 images

Local Binary Pattern(LBP) for extraction of

features,

classification with

(SVM) k-FLBPC

combining LBP with Contour Mask removes noise

51 features are calculated 4 classes are optimized to obtain classification accuracy The accuracy achieved is

98.63%

Similar Types of crops and weeds are classified and real-time weed detection is capable

13 Petrellis[16] 2018 Similarity

techniques between the detection of human and plant diseases have been proposed

Total training samples -

2500

Classes:

Images of Orange fruits and Human skin diseases

Dark spots are

considered using the Octave method

Segmentation and classification using

fuzzy logic MRI scan using Gaussian low pass filter,

classification using

ANN classifier

The accuracy of supervised clustering

is much higher than unsupervised clustering which is

92%

Accuracy remains high on using image processing techniques

14 McDermott

[17]

2011 Prediction of

secreted effectors that reflects biologically relevant features for recognition

Classes:

Type III and Type IV gram-negative bacteria

Classification of

proteins using SVM,

accessing using

SIEVE server Area

of AUC having the

curve under ROC,

Sequence Order

Independence (SOI)

have been calculated

Sequence Order Independence (SOI) worked better The signal of the classes

of the effector has a loosely defined motif

Experimental results found that color information is very much important for disease identification

ROC AUC of Type IV effectors was extremely good SIEVE and Effector had a similar ratio to SOI and SOD

15 Amara [18] 2017 A deep

learning approach that classifies the predictive performance

of unseen diseases of banana leaves

Classes:

1643 healthy images

Images of Banana diseases:

Black

Sigatoka-240, Banana

speckle-1817

LeNet architecture

distinguishing background pixel

using Otsu method, clustering using K-Means have been

used

Precision, F1-score, Accuracy, Recall are combined and the evaluation processed good results RGB format converted as HIS format for better performance

The model provides decision support for the identification of plant diseases by farmers

16 Sperschneider

[19]

2015 Development

of an Effector

in Fungi for conserving features such

as N-terminal sequence motifs

Manually developed secreted proteins –

1922 samples

Predicting Effector candidates

samples

Feature vectors calculated using pepstats, frequencies

of Amino Acids, Molecular length and weight of the sequence, and protein net charge

Feature selection strategy using

Greedy Hill-Climbing Search

and Exhaustive

Experimentally, effectors of 58 fungi were found from species of 16 fungi with a positive sequence set

combination with the expression of planta achieves candidates

of the effector with higher priority

EffectorP helps in predicting species-specific and core effectors Mechanisms for biologically leveled effectors worked well with all species

Search

EffectorP has been trained using

SIGNAL 4.1 as the

initial predictor for

secretomes

17 Al

Bashish[20]

2011 A proposed

method to study, design, implement and evaluate plant leaf disease for its automatic detection and classification

Sample images:

Leaves collected from the area of Al-Ghor present in Jordan

RGB images are converted to color

space, K-Means for

segmentation, features of textures

extracted by Color Co-occurrence Method SGDM

matrices generated for H and S

Extracted features passed through a neural network for recognition

On comparing the 3 models M1, M2, M3,

it was found that M3 emerged to be the best model Neural Network classifier worked well and successfully detects the diseases

Accuracy of precision

93%

Elimination of intensity reduces variations in the intensity

Misclassification occurred for Late Scorch, Cottony

Whiteness, and also for the Normal leaves

18 Sladojevic[21] 2016 Developed a

method to recognize plant leaf disease with DCNN

Total

classes:

15 out of which

diseased leaves – 13

Images for training -

30880 Images for validation

2589

Classes:

Peach, Powdery Mildew, Apple, Grapevine, and Wilt

Detection of plant disease was achieved

by extracting the shape features method

Augmentation for increasing the dataset

overfitting, translation using

Affine transformation,

convolutional and fully connected

layers using CNN and ReLus For

color models

YCbCr, HIS and CIELAB have been

used for the study

10-fold cross-validation technique evaluated a predictive model and repeated it after every thousand training iterations

The network was finely tuned to fit the plant leaves database

Good accuracy achieved after 100th iterations The accuracy achieved is

96.3%

Diseased and healthy

distinguished correctly by the proposed system

19 Ramesh[22] 2020 Proposed a

method for paddy leaf disease recognition and disease classification along with optimized deep neural-based network

Their main aim was to remove the background noise and clustering the diseased portion

Total samples -

650 images

Classes:

Bacterial Blight images:

125, Normal plant images: 95, Blast images:

170, Brown Spot images:

150, Sheath rot images:110

Image background

removed using HUE

values, RGB model converted to HSV model, clustering using K- means done

on the HUE part, the normal and diseased

differentiated using a threshold value

GLCM helps in extracting color cum texture features For iteratively updating the weights of the nodes a combination

of JOA and DNN

has been used

Blast disease achieved higher accuracy when combined with the classifier called

DNN_JOA

Confusion matrices predict the true values

of positive, negative, and false values of positive, negative

DAE, ANN, DNN

compares the results

experimentally found

It was found that the

DNN_JOA classifier

was better than the other classifiers

video detecting architecture for detection

of diseases in crops

Classes:

Rice having stem border images:

1760 images,

Lesion spots with a heavy infestation of diseases have been considered The still-images detection metrics related to True Positive and

It is found that when the learning rate decreases, the iterations also decrease Detection speed limits to 0.1s per frame Precision

The proposed system showed a better result Detection from images causes some problems that include Part Occlusion, Motion Blur, Video

occurred during real-time along with pests

Frame extraction modules, still-image detector, and video synthesizer are part of the system A custom DCNN backbone has been proposed for Faster-RCNN

Rice Brown spot images:

1760 images, Rice of Sheath Blight's images:

1800

Total videos 15,

23, 13

Negative, etc for calculating the lesion spots For videos, it is seen how many boxed lesions are correct They are the

extraction of Frame

module, synthesizer for Video, detectors for Still Images

Detectors with still images work with

ResNet-50, VGG16,

as well as

ResNet-101, and for backbone, the architecture of DCNN was designed

For proposed DCNN,

ReLus layer was put

convolutional layers

Convergence done

using SGD

value using VGG16 and DCNN achieved the highest accuracy

Defocus Lesion spot’s shape is always irregular

21 Mohanty[24] 2016 Development

smartphone-assisted system for disease diagnosis in plants

AlexNet and GoogleNet have been modified

Total samples:

54,306 images

Classes:

14 crop species,

26 diseases

Images collected from the Plant Village dataset

Different sizes of convolution layers

considered

GoogleNet used 9

hyperparameters

such as Stochastic Gradient Descent (SGD) Variations in

color, gray-scale, and segmentation can be seen between

GoogleNet

From the model, the types of crop species and diseases can be found out GoogleNet worked better than AlexNet Accuracy

between 85.53% to 99.34%

The model performed better for color measurement

22 Dey [25] 2016 Design of a

system to detect pests, like White Flies, present

on the plant leaves

Different classifiers were studied

to see the differences between healthy and diseased leaves

Total samples:

200 images

125 affected and

75 normal images

Segmentation is done with K-Means

technique, extraction

of features with the

help of GLCM cum GLRLM The classification was done with the help of

Different classifiers

such as Binary with Decision Tree, Bayesian, K-NN

have been compared

Accuracy with its highest value achieved by SVM, Radial Basis with Function kernel

Performance testing for 5 trained classifiers is done which was not present in the training set Accuracy of

SVM 98.4%

Among the classifiers, the SVM gave the best performance

23 Keh[26] 2020 Investigated

the classification

of pathology problem with single leaf image A new EfficientNet related model with the training for the noisy background was introduced

Total samples:

1820 images of apple leaves

Embossing the image with 0.5 probability, sharpening and applying blur on the images The loss function is used in cross-entropy

EfficientNet and Semi-Supervised Noisy student

training has been compared Adam Optimizer has been

applied

EfficientNet performed better than the other models The weights from the Noisy Student Training achieved good accuracy

Accuracy is higher than other classifiers

EfficientNet and Noisy Student Training when combined yields better results

24 Chohan [27] 2020 Development

of a CNN architecture for detection

of plant diseases

Total samples:

56236 images

Classes:

tomato, apple, raspberry, soybean, squash

Augmentation for increasing the weights of database, extraction of features

architecture, reducing sizes of the

image using Pooling,

and scaling the data

Batch Normalization has

been used

The model classified

a maximum number

of images accurately

Horizontal and vertical edges, RGB values extracted from CNN The model used fully connected layers for prediction

achieved is 95%

CNN was used for correctly predicting the diseases of plant leaf images

25 Minaee[28] 2020 A survey has

been done on the literature review structure on the works of semantic and instance-level segmentation More than

100 deep-learning and segmentation methods that were proposed until 2019 have been discussed

Dataset:

PASCAL VOC 2012,

MS COCO, Cityscapes, and ADE20k

CNN architecture used by researchers

in the survey paper is AlexNet, VGGNet, ResNet, GoogleNet, MobileNet, and DenseNet Encoder decoder models for image translation and sequence to sequence model in NLP, GANs FCN, U-Net, and V-Net for segmentation

Markov, Conditional Random fields, are

architectures RGB segmentation,

convolutional Network, LinkNet, etc have been used

by some Some of them even used augmentation

increasing the number of labeled samples

Through this survey paper, one can learn

segmentation

algorithms proposed till 2019 Different segmentation algorithms have been reviewed including deep learning, training the data,

architectural network, functional loss, strategies for training

20 popular image segmentation dataset has been overviewed

Based on the properties and performance of the reviewed methods, a comparative study has been made

The authors have surveyed more than

segmentation algorithms and grouped them into 10 categories like Adversarial,

Generative, FCN, CNN models, etc

of a new deep learning scheme for masking semantic segmentation models based

on CNN at the pixel level

Total samples: 50

images of Powdery Mildew

Image transformed into HSV color space, extraction of S channel, masking

using Otsu method

For obtaining black background, RGB images were masked

U-Net is used as

CNN architecture

Data augmentation

has been done Adam

method used for optimization For initializing the

weights, the Glorot

method was used

On adding U-Net

normalization, and convolutional kernel, the training process gets accelerated The model achieved satisfactory

segmentation Edges found under the proposed model as well as clustering were smoother than K-Means On higher pixel accuracy, the healthy region predicted accurately

Easy implementation

of the proposed model Shape, as well

as the required area, can be provided for the diseased region In terms of Recall, the model worked better

27 Iqbal [30] 2018 Review of

different methods for detecting diseases and classification

in citrus plants Detail taxonomy has been described

Classes:

Citrus diseases such as Canker, Blackspot, Citrus Scab, Melanoses, Greening, and Anthracnos

e

Challenges such as noisy background, best ROI, changes in the part of a light, low-intensity values, and a huge number of correlated

redundancies have been studied

Different preprocessing techniques,

Efficiency improvement in the terms color space, YCbCr, images, along with K-Means clustering algorithm

Color, texture features using GLCM

classification, a backpropagation neural network was

researchers in their survey used saliency-based techniques which works well for

disease detection FS

and DL methods achieved high accuracy and perfect classification time