comparative study of leaf disease diagnosis system using...

International Journal of Applied Engineering Research, ISSN 0973-4562 Volume 13, Number 19 (2018) pp. 14334–14340© Research India Publications, http://www.ripublication.com

Comparative study of Leaf Disease Diagnosis system usingTexture features and Deep Learning Features

Ashwini T SapkalDepartment of Information Technology, Army Institute of Technology, Pune, India, 411015.

Uday V Kulkarni

Shree Guru Gobind Singji Institute of Engineering and Technology, Nanded, India.

Abstract: The feature extraction technique plays a verycritical and crucial role in automatic leaf disease diagno-sis system. Many different feature extraction techniquesare used by the researchers for leaf disease diagnosiswhich includes colour, shape, texture, HOG, SURF andSIFT features. Recently Deep Learning is giving verypromising results in the field of computer vision. In thismanuscript, two feature extraction techniques are discussedand compared. In first approach, the Gray Level CovarianceMatrix(GLCM) is used which extracts 12 texture featuresfor diagnosis purpose. In second appraoch, the pretraineddeep learning model, Alexnet is used for feature extractionpurpose. There are 1000 features extracted automaticallywith the help of this pretrained model. Here Backpropaga-tion neural network (BPNN) is used for the classificationpurpose. It is observed that the deep learning features aremore dominant as compared to the texture features. It gives93.85% accuracy which is much better than the texture fea-ture extraction technique used here.

AMS subject classification:Keywords: Leaf disease, Deep learning, Alexnet, Texturefeatures.

INTRODUCTION

The agriculture is the main occupation in India which suf-fers with low yield. The main cause of the low yield is thelate diagnosis of the plant diseases. Many times the expertadvice is reached to the farmer very late which generates

need of accurate and fast automatic plant disease detectionsystem. Many researchers have taken efforts to implementautomatic leaf disease detection system [1], [2], [4], [3], [5].In automatic plant disease detection system, the infectedleaf images are provided as input to the system and accuratediagnosis of the disease is the outcome of the system. Thesuccess of this system is dependent on the feature extrac-tion techniques used in it. There are two types of featureextraction techniques: 1) The image processing algorithmsare used to extract features from the infected leaf imagesinput to the system. Colour, Shape, Texture, HOG, SURFand many more such kinds of features are extracted usingimage processing algorithm [6], [7], [8]. 2) The automaticfeature extraction techniques are used which will automat-ically detect features from the input image. Deep Learningis the recent and most advanced technology, which auto-matically extract the features from the given input image[9], [10].

In this manuscript, the automatic leaf disease diagnosissystem is implemented using the above mentioned featureextraction techniques. In the first approach the image pro-cessing based feature extraction technique is used, whichextracts the texture features from the input image. The tex-ture features are extracted using Gray Level CovarianceMatrix. In second approach, the pretrained model of deeplearning, Alexnet is used to extract features from the inputimages. The Alexnet model will automatically extract thefeatures from the given infected leaf image. As Alexnetis the pretrained model, it requires very less amount oftime to extract the features from the given image. Herethe performance analysis of these two feature extractiontechniques is done further. In both the approaches, the


extracted features are input to the Backpropagation neuralnetwork (BPNN) algorithm. The BPNN suffers with theproblem of the slow convergence. But it is observed herethat the BPNN converges faster with deep learning featuresas compared to the texture features.The BPNN requires lessnumber of epochs and its accuracy is also higher, whenit is used with deep learning features. In section 2, thedifferent feature extraction and classification techniquesused in leaf disease diagnosis system are reviewed. In sec-tion 3, the stepwise implementation of the automatic leafdisease detection system with the above mentioned two dif-ferent feature extraction techniques is explained. Section 4,compares the performance of these two feature extractiontechniques. Finally section 5 concludes the paper.

LITERATURE REVIEW ON FEATUREEXTRACTION TECHNIUQES USED INAUTOMATIC LEAF DISEASE DETECTIONSYSTEM

Precision agriculture is the field in which recent technologyis used to improve the productivity of the agriculre field.Many automatic leaf disease detection systems are devel-oped using image processing, computer vision and machinelearning algorithms. In this section the brief survey of thefeature extraction techniques and classification algorithmsused in automatic leaf disease detection system, are car-ried out. In [1] Vijai Singh and A K Mishra present thealgorithm for image segmentation using genetic algorithm.In [11] Zhang Chuanlei et. al. implemented apple diseasedetection system by using 38 features of colour, textureand shape. The most valuable features are identified usingGenetic algorithm technique alongwith correlational basedfeature selection method. Finally, Support Vector Machine(SVM) algorithm is used as a classifier. In [12], SukhvirKaur and Shreelekha Pandey studied and implementedSoyabean disease detection system. They used PlantVil-lage database and combined colour and texture featuresto identify healthy and unhelthy leaves of Soyabean crop.In [13], Shi Yun et. al. used statistical features of colour,texture and shape along with the optimal metrological fea-tures with the highest accuracy rate. They used Probabilitybased Neural Network(PNN) for the classification pur-pose. In [14] author proposed the method in which thecolour and shape features are used together. They also pro-posed new features, mean and peak indices of histogramof shape as a disease propoerty. Total 260 colour based

attributes and 163 shapes based features are searched tofind the best potential features based on differnt aspectssuch as probability of feature error, correlation and targetedclass relevency. Experimental results shows that combin-ing colour and shape features together gives better resultsrather than using them individually. In [15] authors haveimplemented cotton leaf disease detection system using12 texture features and applied on Backpropagation neuralnetwork. In [16] different feature selection techniques areevaluated on turmeric crop dataset. Here PCA, Informa-tion gain and Relief-f attribute methods were investigatedand author claimed that feature selection with Informationgain method with SVM gives better results. In [17] theauthor used Gabor wavelet transform for feature extractionand used SVM classifier in leaf disease detection system.They claimed 89% accuracy. In [18] Histogram of gra-dient (HOG) features are used for leaf disease detectionsystem. In [19], the combination of texture and gradientfeatures are used for plant disease detection purpose. In[20], Zhihua Diao et al. introduced invariant memory the-ory for shape feature extraction and seven Hu invariantmoment parameters were defined as shape features. In [21]author proposed graph cut based approach for the segmen-tation of the cotton disease infected leaves. Recently manyresearchers have proved that deep learning gives a veryprominent results on crop disease identification problem.In [22] S.P. Mohanty used deep learning for plant diseaserecognition. He used Alexnet and Googlenet deep learn-ing model for leaf disease detection. In [23] the maize leafdisease are identified using deep Convolutional Neural Net-work (CNN). In [10] Lenet deep learning model is appliedfor Banana leaf disease detection. In [24] real time deeplearning model is used to implement robust real time tomatodisease. In [4] transfer learning concept of deep learningis used for leaf disease identification. Instead of trainingmodel from scratch, in transfer learning the knowledge ofthe pretrained deep learning model is used. which savesthe time of the network. Here the summary of the litera-ture survey states that the majorly colour, shape and texturefeatures are used for leaf disease detection. Recently deeplearning and transfer learning are used in identification ofleaf diseases. In this article the comparative analysis ofthe traditional feature extraction with recent deep learningbased system is done.

METHODOLOGY

In this manuscript, the analysis of two feature extractiontechniques are done. Here in the first approach the Leaf

14335


Disease Diagnosis system is implemented using the texturefeatures. In the second approach the features are extractedusing pretrained deep learning model, Alexnet. In both theapproaches the Backpropagation neural network algorithmis used as a classifier. Both theses models are applied oncommon small size dataset. Here the pomogranate leafimage dataset is used [25] which has total 73 color images.There are total 5 classes present in database

1. Alternari Alternata2. Anthracnose3. Bacterial Blight4. Cercospora Leaf Spot5. Healthy Leaves

Further the following two appraoches are discussed.

• Leaf Disease Diagnosis system using texture fea-ture extraction technique and Backpropagation neuralnetwork

• Leaf Disease Diagnosis system using Deep learningfeatures and Backpropgation neural network

Leaf Disease Diagnosis system using texture featureextraction technique and Backpropagation neuralnetwork

In this method, 12 texture features are extracted using thegray level covariance matrix (GLCM). These 12 featuresare Contrast, Correlation, Energy, Homogeneity, Mean,Standard Deviation, Entropy, RMS, Variance, Smoothness,Kurtosis and Skew. Below the stepwise algorithm is givenwhich describes the approach 1 in depth.

In this algorithm, Image preprocessing is done by step 2and 3. Image segmentation is done using K-means cluster-ing algorithm. From infected segment the texture featuresare calculated using the GLCM. Futher these features areinput to the backpropagation algorithm and its accuracy iscalculated.

0.1. Leaf Disease Diagnosis using Deep Learningfeatures and BPNN

In the second approach, the automatic feature extractiontechnique is used in leaf disease diagnosis system. Herethe deep learning pretrained model, Alexnet [26] is usedto extract the features from the infected RGB leaf image.In Alexnet 5 convolution layers, 3 fully connected layers

along with dropout is implemented. Below table 1 showsdetail layout of layers used in Alexnet.

Table 1. Alexnet Layers Layout

Sr No Label Layer Description

1 ’ data’ Image Input 227x227x3 images with’zerocenter normalization

2 ’conv1’ Convolution 96 11x11x3 convolutions with stride [4 4] and padding [0 0]

3 ’relu1’ ReLU ReLU

4 ’norm1’ Cross Channel Normalization cross channel normalization with 5 channels per element

5 ’pool1’ Max Pooling 3x3 max pooling with stride [2 2] and padding [0 0]



8 ’norm2’ Cross Channel Normalization cross channel normalization with 5 channels per element









17 ’fc6’ Fully Connected 4096 fully connected layer


19 ’drop6’ Dropout 50% dropout



22 ’drop7’ Dropout 50% dropout


24 ’prob’ Softmax softmax

25 ’output’ Classification Output crossentropyex with ’tench’, ’goldfish’, and 998 other classes

The features are extracted using this pretrained deepneural network, Alexnet. The output of ’fc8’ layer (Layer

14336


No 23 in Table 1) is stored as a feature vector. The back-propagation algorithm is used as a classifier. Here, the sameleaf disease dataset is used for the experimentation purpose.As the input given to the Alexnet model is 227X227X3image, all the leaf images are reshaped to size 227X227.Below algorithm 2 describes stepwise implementation ofthe leaf disease diagnosis system using deep learning fea-tures and BPNN. In this system the Alexnet is used for thefeature extraction purpose. As Alexnet is pretrained model,it requires less amount of time to extract the features. Thenthese extratced features are given as a input to the back-propagation neural network algorithm. Finally the resultsare observed and compared with the first approach.

RESULT ANALYSIS

In this manuscript, leaf disease dignosis system is imple-mented using two different feature extraction methods andBackpropagation neural network algorithm. Here the com-parative analysis of these two feature extraction techniquesis done. In approach 1, the texture feature extraction tech-nique is implemented using GLCM, whereas in approach2 the features are extracted using the Alexnet, pretraineddeep learning model. Both these approaches have used theBPNN as a classifier. Here the performance of the back-propagation is evaluated on these two different featureextraction techniques. The major drawback of the BPNNis its slow convergence rate. Hence the first parameter used

FIGURE 1. Comparative analysis of Average Epochs of 10fold validation.

FIGURE 2. Comparative analysis of Average accuracy of10 fold validation.

for the evalution of these two approaches is BPNN conver-gence time and epochs. In both the approaches, the 10 foldvalidation is used. Hence the second evaluation parame-ter is the average validation accuracy. The performance ofthe BPNN is very sensitive to the initial weight conditions.Hence here the performance parameters are calculated byrunning BPNN 30 times and results are calculated using theaverage epochs required for the convergence and the aver-age of validation accuracy calculated in each of the 30 runs.Figure 1 represents the graph which compares the averagenumber of epochs required by BPNN for convergence pur-pose by using approach 1 and approach 2. In approach 1,the backpropagation algorithm could not converge to itsMSE i.e. 0.001. The BPNN terminates when the numberof epochs reaches to 5000. The graph clearly shows that theapproach 2 requires less number of epochs as compared tothe approach 1.

Figure 2 shows the graph of the average accuracy in10 fold validation, achieved by BPNN, when applied withapproach 1 and approach 2. Below the confusion matrix

14337




created in approaach 1 and approach 2 is displayed in figure3 and 4 respectively.

These results have also proved that the approach 2 ismore outperforming as compared to the approach 1. Basedon the above results, it can be stated that deep learningfeatures used in approach 2 gives very prominent featuresand works much better than the traditional features used in

approach 1. Due to the pretrained model of deep learning,the less amount of convergence time is required for thebackpropagation. The parameters used in backpropagationare given in table 2.

Table 2. Initial value of parameters used in BPNN

Sr No Parameter Inital value Description

1 P 73 Total number of Input Images

2 I 12 in approach 1 Total numbers of nodes in Input layer

1000 in approach 2

3 J 10 Total numbers of nodes in Hidden Layer

4 K 5 Total numbers of nodes in Output Layer

5 η 0.1 Learning Rate

6 MSE 0.001 Mean Squared Error

7 MaxEpochs 5000 Stopping criteria for BPNN

1. Conclusion

In this paper a leaf disease diagnosis system is tested usingtwo different feature extraction techniques on commonsmall size dataset. The texture features and deep learningfeatures are extracted from leaf image dataset and appliedto the Backpropagation neural network for leaf diseasedetection. Deep learning features when applied to the back-porpagtion algorithm gives an average 10 fold validationaccuracy 93.85%. In case of texture features when appliedto BPNN gives an average 10 fold validation accuracy74.83%. Also backpropagation converges to MSE 0.001within MaxEpochs i.e 5000, when applied deep learningfeatures whereas it does not converge to MSE 0.001 within5000 epochs when applied to the BPNN. Hence it is statedthat the deep learning features are better than the texturefeatures. In future the performance of these feature extratc-tion techniques can be evaluated on huge dataset of leafimages available on plantvillage.

REFERENCES

[1] Singh, Vijai, and A. K. Misra. Detection of plant leafdiseases using image segmentation and soft computingtechniques, Information Processing in Agriculture 4, no.1 (2017): 41–49.

[2] Dhingra, Gittaly, Vinay Kumar, and Hem Dutt Joshi.Study of digital image processing techniques for leafdisease detection and classification, Multimedia Toolsand Applications 77, no. 15 (2018): 19951–20000.

14338


[3] Gupta, T. Plant leaf disease analysis using image pro-cessing technique with modified SVM-CS classifier, Int.J. Eng. Manag. Technol 5 (2017): 11–17.

[4] Ramcharan, Amanda, Kelsee Baranowski, PeterMcCloskey, Babuali Ahmed, James Legg, and DavidP. Hughes. Deep learning for image-based cassavadisease detection, Frontiers in plant science 8 (2017):1852.

[5] Johannes, Alexander, Artzai Picon, Aitor Alvarez-Gila, Jone Echazarra, Sergio Rodriguez-Vaamonde, AnaDÃez Navajas, and Amaia Ortiz-Barredo. Automaticplant disease diagnosis using mobile capture devices,applied on a wheat use case, Computers and Electronicsin Agriculture 138 (2017): 200–209.

[6] Velmurugan, P., and M. Renukadevi. Detection ofUnhealthy Region of Plant Leaves and Classificationof Plant Leaf Diseases using Texture Based Cluster-ing Features, Artificial Intelligent Systems and MachineLearning 9, no. 1 (2017): 8–10.

[7] Bakar, MN Abu, A. H. Abdullah, N. Abdul Rahim, H.Yazid, S. N. Misman, and M. J. Masnan. Rice Leaf BlastDisease Detection Using Multi-Level Colour ImageThresholding, Journal of Telecommunication, Electronicand Computer Engineering (JTEC) 10, no. 1-15 (2018):1–6.

[8] Patil, Jayamala Kumar, and Raj Kumar. Analysis ofcontent based image retrieval for plant leaf diseasesusing color, shape and texture features, Engineering inAgriculture, Environment and Food 10, no. 2 (2017):69–78.

[9] Lee, Sue Han, Chee Seng Chan, Simon Joseph Mayo,and Paolo Remagnino. How deep learning extractsand learns leaf features for plant classification, PatternRecognition 71 (2017): 1–13.

[10] Amara, Jihen, Bassem Bouaziz, andAlsayedAlgergawy.A Deep Learning-based Approach for Banana Leaf Dis-eases Classification, In BTW (Workshops), pp. 79–88.2017.

[11] Chuanlei, Zhang, Zhang Shanwen, Yang Jucheng, ShiYancui, and Chen Jia. Apple leaf disease identificationusing genetic algorithm and correlation based featureselection method, International Journal of Agriculturaland Biological Engineering 10, no. 2 (2017): 74–83.

[12] Kaur, Sukhvir, Shreelekha Pandey, and Shivani Goel.Semi-automatic leaf disease detection and classificationsystem for soybean culture, IET Image Processing 12,no. 6 (2018): 1038–1048.

[13] Yun, Shi, Wang Xianfeng, Zhang Shanwen, and ZhangChuanlei. PNN based crop disease recognition with leafimage features and meteorological data, InternationalJournal of Agricultural and Biological Engineering 8,no. 4 (2015): 60–68.

[14] Siricharoen, Punnarai, Bryan Scotney, Philip Morrow,and Gerard Parr. Texture and shape attribute selectionfor plant disease monitoring in a mobile cloud-based

environment, In Image Processing (ICIP), 2016 IEEEInternational Conference on, pp. 489–493. IEEE, 2016.

[15] Ghaiwat, Savita N., and Parul Arora. Cotton leaf diseasedetection by feature extraction, In Research Advances inthe Integration of Big Data and Smart Computing, pp.89–104. IGI Global, 2016.

[16] Pream Sudha, V. Feature Selection Techniques for theClassification of Leaf Diseases in Turmeric.

[17] Prasad, Shitala, Piyush Kumar, Ranjay Hazra, and AjayKumar. Plant leaf disease detection using Gabor wavelettransform. In International Conference on Swarm,Evolutionary, and Memetic Computing, pp. 372–379.Springer, Berlin, Heidelberg, 2012.

[18] J. Jeya Priyankha and K. Suresh Kumar. Crop diseaseidentification using a feature extraction HOG algorithm,Asian Journal of Applied Science and Technology, 1 No3, pp 35–39, April 2017.

[19] Kaur, Ramandeep, and Sanjay Singla. Classification ofPlant Leaf Diseases Using Gradient and Texture Fea-ture. In Proceedings of the International Conference onAdvances in Information Communication Technologyand Computing, pp. 96. ACM, 2016.

[20] Diao, Zhihua, Anping Zheng, and Yuanyuan Wu. Shapefeature extraction of wheat leaf disease based on invari-ant moment theory. In International Conference onComputer and Computing Technologies in Agriculture,pp. 168–173. Springer, Berlin, Heidelberg, 2011.

[21] Prashant R. Rothe, Ravindra. V. Kshirsagar. Automatedfeature extraction in cotton leaf images by graph cutapproach, Research world Press, International Confer-ence, Kuala Lumpur Malaysia, October 2014.

[22] Mohanty, Sharada P., David P. Hughes, and MarcelSalath. Using deep learning for image-based plant dis-ease detection, Frontiers in plant science 7 (2016):1419.

[23] Zhang, Xihai, Yue Qiao, Fanfeng Meng, ChengguoFan, and Mingming Zhang. Identification of Maize LeafDiseases Using Improved Deep Convolutional NeuralNetworks. IEEE Access (2018).

[24] Fuentes,Alvaro, SookYoon, Sang Cheol Kim, and DongSun Park. A robust deep-learning-based detector forreal-time tomato plant diseases and pests recognition,Sensors 17, no. 9 (2017): 2022.

[25] https://in.mathworks.com/matlabcentral/fileexchange/55098-plant-leaf-disease-detection-and-classification-using-multiclass-svm-classifier

[26] Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hin-ton. Imagenet classification with deep convolutionalneural networks. In Advances in neural informationprocessing systems, pp. 1097–1105. 2012.

Ms. A T Sapkal pursued Bachelor of Engineering fromSRTMU Nanded, India in 2002 and Master of Engineer-ing from Pune University in year 2010. She is currently

14339


pursuing Ph.D. from SGGSIET and currently working as anAssistant Professor in Department of Information Technol-ogy. Army Institute of Technology, Savitribai Phule PuneUniversity, India since 2003. She is a member of IEEE since2018, a life member of the ISTE since 2004, CSI mem-ber since 2014. She has published more than 10 researchpapers in reputed international journals and conferencesand it’s also available online. Her main research workfocuses on Neural Network Algorithms, Pattern recogni-tion, and Computational Intelligence. She has 15 years ofteaching experience.

Dr U V Kulkarni has received doctoral degree fromSRTMU, Nanded in 2002. He has total 30 years of teach-ing experience and currently working as a Professor in theDepartment of CSE at SGGSIET, Nanded. His researchinterests include Fuzzy Neural Networks and Pattern Clas-sification. He is a recipient of national level gold medalin the Computer Engineering Division for his researchpaper Fuzzy Hypersphere Neural Network Classifier pub-lished in the journal of Institution of Engineers in 2004and the best paper award for the research paper presentedin international conference held at Imperial College Lon-don, U.K., 2014. He has published more than fifty researchpapers in the field of Neural Networks, Fuzzy Logic andHybrid Computing Systems in the reputed conferences andjournals.

14340

comparative study of leaf disease diagnosis system using...

Documents