8 IX September 2020

https://doi.org/10.22214/ijraset.2020.30031
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.429
Volume 8 Issue IX Sep 2020- Available at www.ijraset.com

A Survey on Pest Detection and Pesticide Recommendation

using CNN Algorithm
Dr. G S Anandha Mala1, Gowtham Kumar K2, Kesavan B3, Kishore R4
1, 2, 3, 4
Department of Computer Science and Engineering, Easwari Engineering College.

Abstract: When we think of crops, we automatically tend to think about the insects because they are the ones which reduces the
crop yield rate and has been a nightmare for farmers. With the technology improvement, we still couldn’t find an efficient way to
sort these issues and the farmers are struggling with the harmful impacts caused to the crops by the variety of insects. The
Present method that the farmers are following for separating the insects from the crops is with the help of man power. But this
requires a lot of man power and it also requires a lot of time when there is a huge crop field. This work makes use of
convolutional neural network model to identify and classify the insects.
Compared with previous classifiers such as k-nearest neighbors and linear discriminate analysis (LDA), support vector machine
(SVM) was proposed with Haar-like features to classify insects and obtained a poor performance than the Convolutional Neural
Network .The proposed model is an Android app and thus it helps effectively for farmers.
Keywords: Convolution neural networks, android app.

I. INTRODUCTION
Convolutional Neural Networks (ConvNets or CNNs) are a category of Neural Networks that have proven very effective in areas
such as image recognition and classification. CNN have been successful in identifying faces and it is done by taking the image, pass
it through a series of convolutional, nonlinear, pooling (downsampling), and fully connected layers, and get an output. The output
can be a single class or a probability of classes that best describes the image. The pre-processing required in a ConvNet is much
lower as compared to other classification algorithms. While in primitive methods filters are hand-engineered, with enough training,
ConvNets have the ability to learn these filters/ characteristics. The architecture of a ConvNet is analogous to that of the
connectivity pattern of Neurons in the Human Brain and was inspired by the organization of the Visual Cortex.
A ConvNet is able to successfully capture the Spatial and Temporal dependencies in an image through the application of relevant
filters. The architecture performs a better fitting to the image dataset due to the reduction in the number of parameters involved and
reusability of weights. In other words, the network can be trained to understand the sophistication of the image better. Here inorder
to help farmers from pests , we use CNN to capture the insects in the field . From the image captured it processes the data with the
samples and gives an output. With that output we suggest the right pesticide for the farmer.

II. LITERATURE SURVEY

In paper[ 1],the authors Lingqiao Liu, Chunhua Shen, Anton Van Hengel discussed about cross convolutional pooling and 2 types of
methods to create image representation. The first method is ,the image is directly feeded into DCNN. The second method is,
applying the DCNN to the input image to subregions . DCNNs is very much helpful in extracting local features of the input image.
The issue is While using k value for pooling method and when k becomes too large, retrieval performance will start to drop.
In paper[2],the authors Rahul Chauhan, Kamal Kumar, R.S.Joshi discussed the as Relu non linearity activation function that has
been used for training neural network models . Relu activation function has been widely used to train larger neural networks and it
is also relatively easier activation function when compared to sigmoid, tanh, and leaky activation functions. The issue is, In case of
high bias which results in underfitting ,the network needs to be trained longer Variance error will occur and it affects the resultant
value. In paper[3],the authors Ann-Katrin Fattal, , Michelle Karg, Christian Scharfenberger, and Jürgen Adamy discussed about the
region proposal networks . The RPN is used as slider over the feature maps of a convolutional layer, then it computes each anchor
scores by using bounding box regression at each different level. Small ZF-Net is used that consists of five convolutional layers, two
pooling and two fully connected layers. The two pooling layers makes sure the input image is not down-sampled, so small objects is
detected. The issue is prior maps which are relying on Visual Attention might produce inconsistent results.
In paper [4] Ke Gu discussed about smoke detection. They used a new deep dual-channel neural network (DCNN) which is made
by selective-based batch normalization network(SBNN) and skip connection-based neural network(SCNN) for feature fusion and
feature extraction for classification of smokes.

©IJRASET: All Rights are Reserved 193

 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.429
Volume 8 Issue IX Sep 2020- Available at www.ijraset.com

So the DCNN consist of two layers the first is multiple convolutional layer and max pooling layer which is used to extract the detail
information of smoke and the second layer is the batch normalization layer which is used to improve the feature propagation. These
two layers can be used in our project for feature identification.
In paper [5] Yang Ji discussed about automatic classification of spider image in natural background based on Convolutional Neural
Network and transfer learning. So here they used inspection v-3 and feature vector for transfer learning model which is used to solve
domain problems with existing knowledge. The image pro processing is done with contour detection and data expansion. And CNN
is used to extract effective feature of spiders from complex images with natural background.
In paper [6], Shaoqing discussed about real time object detection using faster R-CNN. The R-CNN is a classifier that identify the
regions on image. It mainly consit of two modules the deep fully convolutional network which is used to propose regions and the
second model which is faster r-cnn that detects the proposed regions. Here the rcnn uses a shared convolutional layer to get
rectangular object proposal. It is an unified, deep learning based object detection that runs at 6-17 frame per seconds.
In paper [7], Sangdi Lin and George C.Runger discussed about a new end to end deep neural network model for time-series
classification ( TSC) with emphasis on both the accuracy and the interpretation. This model consists of a convolutional network
component to extract high level features and a recurrent network component to enhance modeling of the temporal Characteristics of
TS data . It also uses sparse group lasso (SGL) to generate final classification. This model gives good interpretability through the
SGL.It outperforms traditional CNN.
In paper [8] , Fok Hing Chi Tivive and Abdesselam Bouzerdoum discussed about some efficient training algorithms based on first
order , second order and conjugate gradient optimization methods. This is a hybrid method derived from the principles of quickprop,
rprop and LS . Alll these methods are combined to give a better recognition pattern of the image .
In paper [9], R Amog Shetty Rishab F TatedSunku Rohan Triveni S Pujar use CNN and train a neural network model that predicts
whether the crop is going to get any pest and disease attacks. The model developed by these authors gave up to 99% classification
ability which satisfies the required efficiency as Neural Network is used. It highly depends on the colour and shape features that are
extracted from the input image the prediction occurs
In paper [ 10] ,Hoo- Chang Shin , Holger R Roth, Mingchen Gao , Le Lu , Ziyue Xu , Isabella Nogues , Jianhua Yao , Daniel
Mollura , Ronald M summers discussed about factors of employing deep convolutional neural networks to computer aided detection
problems Here the different CNN architectures are explored and evaluated .It contains 5 thousand to 160 million parameters and
finally the examination of transfer learning is done .This is more efficient than normal CNN methods.

Table1.1 CNN in different fields

S.NO PAPER TECHNIQUE RESULT ISSUES
1. Convolutional layer Pooling Cross-layer pooling In Comparison Cross- layer While using k value for pooling
for Image Recognition pooling counterpart achieves method and when k becomes too
74.4% on MIT67 which is large,retrieval performance will start
much better than the traditional to drop.
pooling
2. Convolutional Neural Relu non linearity The calculated accuracy on In case of high bias, which is on
Network(CNN) for Image activation function MNIST is 99.6% . training batch the network needs to
Detection and Recognition be trained longer .
Variance error will occur
3. Saliency-Guided Region RegionProposed The region proposal network The issue is prior maps which are
Proposal Network for CNN Network,Small ZF-Net increases the overall execution relying on Visual Attention might
Based Object Detection time of up to 3ms . produce inconsistent results
4. Deep Dual-Channel Neural DCNN and SBNN SCNN The DCN has achieved high the DCNN fails with the smoke that
Network for Image-Based performance on average as has fewer textures.
Smoke Detection compare to HLTMPC.

5. Automatic Classification of Convolutional Neural the accuracy of training set and The accuracy and loss of training set
Spider Images in Natural Network, Transfer testing set can reach more then changed with time and CNN is too
Background Learning, Contour 90% recognition speed can be simple and it was not able to extract
detection and data controlled with 1 second. accurate feature of spiders.
expansion for
preprocessing of image.

©IJRASET: All Rights are Reserved 194

 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.429
Volume 8 Issue IX Sep 2020- Available at www.ijraset.com

6. Faster R-CNN: Towards ROI Pooling , It detects objects of wide range The system failed in high noises area.
Real-Time Object Detection Convolutinal layers were of scales and aspect ratio.it
with Region Proposal used for object region require only 198ms to identify
Networks identification. regions which is faster the
normal rpn.
7. GCRNN: Group-Constrained Sparse group lasso , CNN The output is satisfying classification accuracy is not large
Convolutional Recurrent classification and also good
Neural interpretability.
8. Efficient training algorithms CNN , first order ,second error rates of lesss than 3% : it takes more time in training the
for a class of shunting order training methods across all architectures data sets
inhibitory convolutional and conjugate gradient
neural networks optimization methods
9. CNN based Leaf Disease CNN, leaf disease, The network is trained with 99.32% is achieved yet it can be
Identification and Classification, deep 99% classification ability with generalized to classify for other
Remedy Recommendation learning, remedies CNN to provide remedy diseases.
System
10. Deep Convolutional Neural pre trained images ,CNN High performance for medical It needs collection of more data sets
Networks for imagining tasks and takes more time
Computer-Aided Detection:
CNN Architectures,
Dataset Characteristics and
Transfer Learning

III. CONCLUSIONS
CNN can be used to give high accuracy in detection and classification of insects compared to other algorithms. A variety of insects
detected and thus generalizing for a number of diseases. The insects images provided as input can be in any of the classes of pest
images included in the training set. The model trained as a result can classify a number of diseases by CNN to give high accuracy
unlike other algorithms as high upto 90%.

REFERENCES
[1] Liu, L., Shen, C., & Hengel, A. van den. (2017). Cross-Convolutional-Layer Pooling for Image Recognition. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 39(11), 2305–2313.
[2] Chauhan, R., Ghanshala, K. K., & Joshi, R. . (2018). Convolutional Neural Network (CNN) for Image Detection and Recognition. 2018 First International
Conference on Secure Cyber Computing and Communication (ICSCCC).
[3] Fattal, A.-K., Karg, M., Scharfenberger, C., & Adamy, J. (2017). Saliency-guided region proposal network for CNN based object detection. 2017 IEEE 20th
International Conference on Intelligent Transportation Systems (ITSC).
[4] K. Gu, Z. Xia, J. Qiao and W. Lin, "Deep Dual-Channel Neural Network for Image-Based Smoke Detection," in IEEE Transactions on Multimedia, vol. 22, no.
2, pp. 311-323, Feb. 2020.
[5] Y. Jian, S. Peng, L. Zhenpeng, Z. Yu, Z. Chenggui and Y. Zizhong, "Automatic Classification of Spider Images in Natural Background," 2019 IEEE 4th
International Conference on Signal and Image Processing (ICSIP), Wuxi, China, 2019, pp. 158-164
[6] Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun
[7] Lin, S., & Runger, G. C. (2017). GCRNN: Group-Constrained Convolutional Recurrent Neural Network. IEEE Transactions on Neural Networks and Learning
Systems.
[8] Tivive, F. H. C., & Bouzerdoum, A. (2005). Efficient Training Algorithms for a Class of Shunting Inhibitory Convolutional Neural Networks. IEEE
Transactions on Neural Networks,
[9] Suma V R Amog Shetty, Rishab F Tated, Sunku Rohan and Triveni S Pujar “CNN based Leaf Disease Identification and Remedy Recommendation System” ,
Proceedings of the Third International Conference on Electronics Communication and Aerospace Technology [ICECA 2019] IEEE Conference Record , Vol.
,pp:395-399
[10] H. Shin et al., "Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning," in
IEEE Transactions on Medical Imaging, vol. 35, no. 5, pp. 1285-1298, May 2016.

©IJRASET: All Rights are Reserved 195