RICE LEAF DISEASES CLASSIFICATION USING CNN WITH TRANSFER LEARNING

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p-ISSN: 2395-0072

RICE LEAF DISEASES CLASSIFICATION USING CNN WITH TRANSFER LEARNING

2Assistant Professor, department of computer application Madanapalle Institute of Technology and Science, Angallu, AP, India ***

Abstract - Riceis one of the maximum critical plants in India and is liable to diverse illnesses at some point of extraordinary tiers ofcultivation.It could be very tough forfarmerswith confined understanding to as it should be perceiving those illnesses manually. Recent traits in deep gaining knowledge of have proven that automated picture reputation structures the use of convolutional neuralnetwork(CNN) fashions arevery beneficial forsuch problems. Since rice leaf ailment picture datasets aren't quite simply available, we created our very own small dataset.Therefore,I advanced adeep gaining knowledge of version the use of switch gaining knowledge of.The proposed CNN structure is primarily based totally on VGG-sixteen and is educated and examined the use of paddy subject and net datasets. The accuracy of the proposed version is92.46%.IndexTerms–Convolutional Neural Networks, Deep Learning, Fine Tuning, Rice Leaf Disease,TransferLearning.

Key Words: ConvolutionalNeuralNetwork,DeepLearning, Fine-Tuning,RiceLeafDiseases,TransferLearning.

1. INTRODUCTION

Rice is the staple delivery of food in India similarly to at some point of the world. It is attacked with the useful resource of the use of hundreds of ailments in numerous levels of its cultivation. Therefore, early detection and remedy of such ailments are beneficial to make certain immoderate quantity and quality quality, but this is very difficultduetothebigexpanseoflandunderneathindividual farmers and the big kind of ailments similarly to the prevalence of more than one disorder withinside the identicalplant.Agriculturalexpertknowledgeisn`tonhand in a way off areas, and its miles a time taking process. Therefore,theAutomatedSystemsarerequired.Toaidthe plightofthefarmersandprovideadvancedaccuracyofplant disorder detection, research artwork using numerous gadgetslearningalgorithmssuchasSupportVectorMachine (SVM) [1]– [3], Artificial Neural Networks [4] have been done.

However,theaccuracyofsuchsystemsishighlyrelyingon feature preference techniques. Recent researches on convolutionalneuralnetworkshaveprovidedtremendous bounce ahead in photograph based totally absolutely

recognition thru removing the need for photograph preprocessingsimilarlytoimpartingconstructedinfunction selection.Anotherchallengeisthatit`smilesverydifficultto gain huge sized dataset for such problems. For times whereindurationofthedatasetisspecificallysmall,it'sfar more main to use a model this is pretrained on a huge dataset. This is called Transfer Learning and it can be implementedtocreateamodelthatcanbeusedasadifficult and speedy function extractor disposing of the very last honestly associated layer or thru fine-tuning the last few layersaverygoodmannertoartwork moreuniquetothe involved dataset. Nowadays, mobileular phones are accessibletoeverybodyandsowe`vegivenyoutheideaof an automated device wherein the farmers can upload the diseased leaf image and post it to our server wherein the neuralnetworkmaybeusedtofindoutthesicknessandthe sicknessmagnificence on theaspectof the remedy can be dispatchedoncemoretothefarmer.Inthoseartwork we havegotproposedtheshapeforthesicknessmagnificence partoftheautomateddevice.Inspiredthrutheartworkin [5]–[8]and[14]onconvolutionalneuralnetworks,onthose artwork,wehavegotsuperiorthedeepmasteringtechnique onourricesicknessdatasetthatwehavegotaccruedover past several months. We have used the pre-professional VGG-16version(TrainedonthemassiveImageNetdata)and the use of Transfer Learning we have got finetuned the actually associated layers just so we're capable of accommodateourvery owndatasetandatthepreventwe have got completed some errors assessment and tried to provideancauseofthereasonsfortheerrors.

2. LITERATURE SURVEY

A lot of research were completed using traditional classifiers but the effects are relying at the characteristic desire techniques and photo preprocessingisa high step [9].Therefore,CNN hasattracted multiple researchers to take advantage of immoderate reputation accuracy

A. Plant Disease Detection using CNN

Alotofresearchhadbeenperformedtheusageoftraditional classifiers but the effects are relying at the characteristic preference techniques and photo preprocessing is a high step [9]. Therefore, CNN has attracted more than one

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p-ISSN: 2395-0072

researcher to take benefit of immoderate recognition accuracy.

B. Rice Disease Detection using CNN

Convolutionalneuralnetworkclassifierisusedonadataset of227photographsofsnail-bitten,diseasedandwholesome riceflowersin[8].Theclassifieristransfergettingtoknow baseddefinitelyusingAlexNet.Trainingtheaboveshapean accuracy of 91.23% is achieved but it can maximum efficientlyanticipatewhetherornotornownolongerplant is diseased or now no longer. In [13], the authors accumulated500photographsof10superbriceillnessesof leaf and stem. They developed a shape inspired with the useful resource of the usage of Le-Net and Alex Net and achieved95.48%onthetestset.Sincethefactscanbeverya lot less, they used several preprocessing steps like image resizingto512*512,normalization,PCAandwhitening.They usedstochasticpoolinginlocationofmaxpoolingandstated thatitpreventsoverfitting.

3. PROPOSED METHOD

Inproposedsystem,weadviseaDeepLearninggeneration thatautomaticallyapprehendpicsusingConvolutionNeural Network (CNN) models can be very beneficial in such problems. By using the ones techniques, we are able to resultseasilydiscoverandselectoutoutthediseases

weregatheredfromtheInternationalRiceResearchInstitute (IRRI) Rice Knowledge Bank website. There were constrainedhugesortofimagesforschoolingoursystem,so we`ve were given used a few statistics augmentation techniqueswiththehelpofKera`sDocumentationtogeta substantialhugesortofimages.Thedatasetconsistsof1649 imagesofdiseasedleavesofricewhichincorporatesthree mostcommonlocationillnessesmainlyRiceLeafBlast,Rice Leaf Blight, and Brown Spot. There are 507 images of Healthyleaves.Wehavenownotaccomplishedanystepto dispose of noise from the raw statistics. There were a number of issues faced at the same time as amassing the statisticslikehorribleilluminationandmorethanoneillness withinsidethesameplant.Wehavetriedtoconquerthem thru manner of method of the usage of picture graph preprocessingstepslikeresizingandzooming.Thehugesort ofimagesthatcouldbegatheredfromthefieldsareverya wholelotmuchlessforschoolingCNNsowehavegotwere givenusedanumberofaugmentationtechniqueslikezoom, horizontalandverticalshift,androtationwhichisprobably stated withinside the Implementation Section later. The beneathNeathsectionsdescribethecommandsofRiceLeaf illnessesonwhichwehavegotweregivenworked.

Fig.1.Overviewofthestepsoftheproposedmodel

4. RICE DISEASE TYPES AND DATASET DESCRIPTION

Thericeimagegraphdatasethasbeenamassedthroughout thepreviousfewmonthswithinsidetheprincipalfromthe cultivation fields of Maharat village (District: South 24 Parganas) in Baruipur, Dharana village District: Pura Medinipur) in Tamluk and Basi hat (District North 24 Parganas),belongingtothedominionofWestBengal,India similarlytofromtheInternet.Thepicsweretakentheusage ofMotorola E4Plus andRedmi5A molecularcamera.The symptomsandsymptomsandstatisticsabouttheillnesses

Fig.2.TypesofRiceLeafDisease

A. Leaf Blast

It is a fungal illness because of Magnaporthe Oryza. The initial symptoms and symptoms and signs are white to grey-green spots which might be elliptical or spindlefashioned with dark pink tobrownishborders.Somehave diamond shape with extensive centers andpointedends.In the Figure 2 (a) the spindle fashioned lesionswith white spotsanddarkbrownbordercanbeseen.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p-ISSN: 2395-0072

networks whose shape determines the general overall performance of the network. It consists of three additives namely, convolution layer, pooling layer and absolutely associatedlayer.Thefirsttogetherofficeworksthefeature extractorandthe0.33-layeractsasaclassifier

•Newdatasetissmallandacquaintedtoauthenticdataset.

•Newdatasetisbigandacquaintedtotheauthenticdataset.

•Newdatasetissmallhowevervariedtoauthenticdataset

•Newdatasetisbigandvariedtoauthenticdataset.

Fig.3.VGG-16Architecturefine–tunedwiththelasttwo layerswith128DenseFCLayerand4DenseSoftMax Layerastheoutput.

b. Leaf Blight

Itisafungal illness because of Magnaportheoryzae.The initial symptoms and symptoms and signs and symptoms arewhitetogrey-green spots which might be ellipticalor spindle-shaped with dark crimson to brownish borders. Some have diamond shape with extensive centers and pointedends.IntheFigure2(a)thespindle shaped lesions

B. Brown Spot

It is a fungal disease. The inflamed leaves have several massive spots at the leaves that can killthe entire leaf.At the preliminary stage, small, round, darkish brown to purple-brown lesions may be located withinside the leaves.Fully evolved lesionsare round tooval with mild brownto grey center,surrounded with the aid of using a reddish-brownmargin because of thetoxinproduced with the aid of using the fungi are the small darkish brown lesionsoftheBrownSpotaffectedleaves

5. METHODOLOGY

Convolutional neural networks (CNNs) are multi-layered networks whose shape determines the general overall performance ofthenetwork.It consists of three additives namely, convolution layer, pooling layer and absolutely associated layer. The first together office works the feature extractor and the 0.33-layer acts as a classifier Convolutional neural networks (CNNs) are multi-layered

Sincewe`vesmalldatasetanditisuniquefromtheImageNet dataset,ourmodelfallswithinsidethe1/threescenario,so we have got steady the layers of the VGGNet to use it as feature extractor until the ultimate absolutely associated layers which we have got fine-tuned in keeping with the styleofcommandsinourdataset.InFigure2theshapeof proposedmodelisdepicted.Wehavemoreoversuperiora CNN model without transfer learning with 4 Convolution layers each of it's determined via ReLU, Maxpooling and dropoutlayerdeterminedvia2FullyConnectedLayerand SoftMax.Butthegeneraloverallperformancewasnownow not as nicely due to the fact the above-said model. The assessmentoftheidenticalhasbeensaidwithinsidethestop endsresultsection

6. IMPLEMENTATION

A. Experimental Setup

ThetakealookatcometobeachievedonaWindows10PC organizedwithGPUcardP4000,64-bit OperatingSystem. TheCNN-basedcompletelymodelcometobeaccomplished withinsidetheKera`s2.2.4deepanalyzingframeworkwith TensorFlow1.13.1backendandpython3.7.2

b. Image Acquisition

The pictures are collected from the cultivation fields similarly to from internet. As referred to withinside the dataset description, statistics encompass 4 schooling mainly Leaf Blast, Leaf Blight, Brown Spot and healthful plant pictures

c Image Preprocessing and Augmentation

The images amassed are resized to 224*224 pixel and a number of augmentation techniques likezoom, rotation, horizontalandverticalshiftare achieved using ImageData GeneratorinKerastogeneratenew images.

D. CNN Model Training

Thepictureinformationsetisloadedfortheeducationand checking out. The elegance labels and the corresponding

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p-ISSN: 2395-0072

photos are saved in respective arrays for education. 70 percentage of information is used for education and 30 percentageofinformationisusedforcheckingouttheusage of teach take a look at cut up feature. The 70-percentage informationissimilarlycutupand20%ofitsmilesusedfor validation.Theelegancelabelsareencodedasintegersand then,one-warmencodingisexecutedonthoselabelsmaking every label represented as a vector in place of an integer. Next,theVGG-sixteenversionisloadedfromkerasandthe remainingabsolutelyrelatedlayersareremoved.Theclosing layersaremadenon-trainable.Wehaveflattenedtheoutput of characteristic extractor part, accompanied through absolutelyrelatedlayerandoutputlayerwithSoftMax.Then we've got compiled our version the usage of the Adam optimizer with categorical_crossentropy because the loss featurefortype.Wehavestoppedat25epochsconsidering after this the consequences had been stable. Figure three suggeststhestairswe'vegotfinishedforthetypeprocess.

E. Justification for the Chosen Model

Transfergettingtoknowrefersbacktothestateof affairs wherein what has been discovered in a single placing is exploitedtoenhancegeneralizationinsomeotherplacing. Transfergettingtoknowhastheadvantageofreducingthe education time for a neural community version and as a consequence could be very beneficial considering the fact that maximum real-global troubles normally do now no longer have hundreds of thousands of categorized information factors to teach such complicated models. Usually,anumberofinformationiswantedtoteachaneural communityfromscratchhowevergetadmissiontotothat informationisn`tconstantlyavailable.Withswitchgettingto knowastrongdevicegettingtorecognizemodelcanbebuilt withquitelittleschoolingrecordsbecauseoftherealitythe model is already pre-professional. Hence, we`ve used the pre-professional VGGNet and fined tuned it to classify the useofournon-publicsmalldataset.

7. RESULTS

A. Calculations

Theproposedversioniscompletedfor25epochsover1509 education informationobserved withtheaid of using 647 checkinformationandtheaccuracyofeducationsetis97% andthecheckaccuracyof92.four%.Wehaveadditionally completedtheequalinformationtheuseoftheequalcutup ratiointotrain,validationandchecksetoneveryotherCNN versionwithoutswitchmastering.Thebatchsize,varietyof epochs,optimizer becamefine-tuned and16, 30,rmsprop respectively together with dropout 0. four supplied the exceptionalendresultbuttheexceptionalaccuracybecame 74%. The CNN version without switch mastering has four Convolutionlayerseveryofthatisobservedwiththeaidof using ReLU, Maxpooling and dropout layer observed with

theaidofusing2FullyConnectedLayerandSoftMax.TableI in Figure four suggests the evaluation in accuracy of the proposed CNN version with Transfer Learning and CNN without Transfer Learning. Figure five illustrates the Trainingandvalidationaccuracyasopposedtothevarietyof epochsfortheCNNwithTransferLearning.

TABLE I. PERFORMANCE OF COMPARISON OF CNN WITH AND WITHOUT TRANSFER LEARNING

Fig.4.PerformancecomparisonofCNNmodelwithand withoutTransferLearning

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p-ISSN: 2395-0072

Fig.6. Performance Comparison of CNN model with and withoutTransferLearning

Fig.6.(a).(b).Comparisonofdiseasesinriceleaf prediction

B. Error Analysis

The Figure 6 (a)-(f) illustrates pictures which are misclassified with the aid of using the proposed CNN model. The misclassifications are defined in information withinside the underneath segment for every of the sickness type.RiceBlast:Image(a)belongstoRiceBlastbut (a) is classed as Brown Spot because the photograph is blurred.The purpose may be thepresenceofsmallbrown spots withinside the identical riceleaf.

Leaf Blight:Images(d)and(e)arecategorizedasHealthy howeverthey belongtoBlightcategory. Thecause will be badilluminationandblurringofimage.

Healthy:Image(f)is healthful however it's far labeled as BrownSpot probable due to the fact the picture isblurred and comparison ispoor

Brown Spot: Images (b) and (c) belong to Brown Spot however are categorized asBlast.One purpose will be the presenceofsmallblastlesions at the leaf.In(d)thebrown spotlesionsresembletheblastlesion

Fig.6.Fromlefttoright(a)-(f)Ricediseaseimagesthat aremisclassifiedbythemodel.(a)Rice Blastdisease(b) and(c)BrownSpot(d)and(e)LeafBlight(f)Healthy

3. CONCLUSIONS

Inthispaperwe'vegotproposedadeepgainingknowledge ofstructurewitheducationon1509picturesofriceleaves and trying out on one of a kind 647 pictures and that successfullyclassifies92.46%ofthetakealookatpictures. Transfer Learning the use of fine-tuning the predefined VGGNethassignificantlyadvancedtheoverallperformance of the version which in any other case did now no longer producehigh-qualityconsequencesonsuchsmall dataset. Thevarietyofepochsusedturnedintostoppedat25dueto the fact we had acquired a reduce factor and then the accuracyturnedintonownolongerimproving,andtheloss turnedintonownolongerreducingoneacheducationand validation data. In destiny work, We would really like to accumulate more pics from agricultural fields and AgriculturalResearchinstitutes simply so we are able to decorate the accuracy further. We would really like to characteristic cross-validation approach in future a great manner to validate our consequences. We could additionally like to apply higher deep getting to know fashions and different state-of the artwork works and examine it with the outcomes obtained. The evolved version may be utilized in destiny to locate different plantleafdiseases, that are crucial vegetation inIndia.