INTERNATIONAL JOURNAL OF RESEARCH IN TECHNOLOGY AND MANAGEMENT (IJRTM)

ISSN 2454-6240
www.ijrtm.com

RICE CROP DISEASE DETECTION USING YOLO ALGORITHM

M. Sandhya
G. Nihar V. Raghavendra
Department of Electronics and
Department of Electronics and Department of Electronics and
Communication Engineering
Communication Engineering Communication Engineering
Sreenidhi Institute of Science and
Sreenidhi Institute of Science and Sreenidhi Institute of Science and
Technology
Technology Technology
Hyderabad, India
Hyderabad, India Hyderabad, India
[email protected]
[email protected] [email protected]
V. Suresh
Department of Electronics and
Communication Engineering
Sreenidhi Institute of Science and
Technology
Hyderabad, India
[email protected]

Abstract—There are many hurdles in trying to produce a Countries that grow rice, have developed their own
successful yield from the crops. Crops are eventually infested programs to increase development of rice crops. For
by pests or infected by diseases which destroys a significant example, in Bangladesh, they promised to increase the
amount of the crop if not detected in the early stages.Rice is production of rice significantly by the introduction of green
one of the major food crops of India. The majority of the
revolution package [7]. A training program that provides
population consume rice as their daily meal. But every year, a
significant part of the crop is destroyed by diseases. Therefore, science-based information to farmers about integrated pest
detection and prevention of such diseases in advance will help and disease management was founded by Farmer Schools in
us have a greater yield and healthy crop. According to the Thailand. This helps the farmers to maximize the yield
statistics given by International Rice Research Institute, about while reducing the diseases [8]. Due to the public policy,
37% of the crop is lost to diseases annually. Modern technology some of the developed countries have neglected the farmers
has advanced enough where farmers are able to make out the and their need for education of the evolving agricultural
diseases from the symptoms shown by the crop and searching science [9]. Though chemicals play a key role in abstaining
for them in the internet through smartphones. So here, we the diseases [10], they are also likely polluting the
propose a method to detect the diseases with the help of YOLO
environment. The chemicals should only be used when there
object detection algorithm. A thorough research was
conducted into the mechanisms of farming and the type of are really needed so that the drawbacks of their usage can be
conditions the crops were grown in, along with the prevention reduced [12]. IPM (Integrated Pest Management) is a
methods of diseases. With the help of image processing and framework that helps the agriculture practitioners to detect
supervised learning, our system detects the diseases so that the diseases and pests and to make tough decisions by
necessary steps can be taken towards the cure. A number of providing them with the scientific knowledgeand technology
images were taken and pre-processed before giving them to the [11]. The in-time detection of lethal diseases can be made
model for training. We were able to achieve an accuracy of possible to the farmers with the help of technology [13].
98.92 percent. This model can be used to help the farmers to Advance computing technologies can help farmers make
detect the diseases and curb the disease in its early stages so
decisions on many aspects of growing crops. Proper
that the yield is maximized at the end of the year.
diagnosis of the crop in fields is a key point in maximizing
Keywords—deep learning, AI, supervised learning, the yield [14]. [13]. Unfortunately, the amount of
classification, cnn, yolo applications of technology is by far less in the field of
agriculture compared to other fields like engineering,
I. INTRODUCTION medicine etc.Due to this, it is still a big problem for the
agricultural society to develop a system for the detection of
Rice is cultivated in more than 100 countries with diseases and calculating the percentage of loss in yield in a
varying environmental conditions [1]. It is a major food crop timely manner [15]. For most cases human visually
in majority of the Asian countries [2]. Unfortunately, there conducts the treatment of diseases. Trained examiners may
are a number of diseases which can weaken the crop and be effective in identifying plant diseases, but some related
damage the yield if not treated on time. Thus, it is important drawbacks may be detrimental to disease recognition efforts
that these diseases are cured in time [3]. Rice crop is [3]. Application of advanced technological tools in the field
destroyed not only by diseases but several pests as well [4]. of agriculture can greatly help the farmers to detect and
Of the 266 species of insects found in the rice fields, more prevent diseases without waiting for an expert’s aid [16].
than 40 of them are pests [5]. Therefore, the need for the Computer vision along with deep learning and image
timely detection of the disease or pest becomes the utmost processing has been applied in various fields of science
priority[6]. which have proven to have applications even in the field of
the today’s evolving agriculture [17]. Machine vision
National Conference On Advances in Electronics Signal Processing and Communications (AESPC-2020)
Volume 6 Issue 3, JUNE 2020
 INTERNATIONAL JOURNAL OF RESEARCH IN TECHNOLOGY AND MANAGEMENT (IJRTM)
ISSN 2454-6240
www.ijrtm.com
techniques are used in addition to the computer vision SoftMax activation function is used at the output layer
techniques which are playing a key role in the protection of which classifies the objects based on the class probabilities.
crops [18]. Further addition of digital image processing has A loss function(lossfunc) is also associated with the network
led to the idea of crop-field management [19]. The use of to calculate the loss/error(errcalc) during training. Generally,
modern computer vision and machine vision techniques categorical_crossentropy function is used to calculate the
helps us to implement a disease detection system which not loss associated with the training of CNN classification
only reduces the amount of time and energy but also models.
performs with higher accuracy[20]. Let’s say there is an image(img) of dimensions v×v (vi),
Many object detection and classification models which is being given to the CNN network. A kernel(kern) of
have been used for the timely detection and control of the dimensions p×p (pi) is being used with valid-padding (no
diseases [21]. In Kawasaki et Research. Al., a CNN based padding). A max pooling layer of size c x c is being used
disease classification system was implemented which with a stride (str) of 1. The output of the given convolution
achieved high classification efficiency by training images layer(convout) will be a feature map of the order (v-p+1, v-
alone [22]. Sladojevic et. in research. Al., a deep p+1). This is again sent to a max pooling layer to extract the
convolutional network was used to build a plant disease more distinguished features which gives an output with the
recognition model based on the classification of the leaf dimensions ((v-p+1)/c, (v-p+1/c)). These are again given to
image. In the field of image classification [23], a convolution layer and the process is repeated for as many
Convolutional neural networks (CNN) achieved amazing convolution layers in the network. In this way each
result. Yao and. In their research, Al. suggested a model for convolution layer extracts a specific feature of the image. As
the detection and recognition of White-backedplanthoppers the number of convolution layers increases so does the
using a three-layered approach. This approach has been feature extracting ability of the network. The output of the
proven to be successful in identifying various planthoppers CNN(Cnnout) is the class probabilities and the class with
developmental stages at rice plants [24]. highest probability is predicted as the output.
As we can see, many researches have conducted by
various authors on the problem statement. Next, we go
through the basic architectures and the steps involved at 𝐶𝑛𝑛𝑜𝑢𝑡 = 𝑖𝑚𝑔 ∗ 𝑓𝑒𝑥𝑡 = ∑𝑁
𝑖=0 𝑤𝑔𝑡𝑖 ∗ 𝑖𝑚𝑔𝑖 ∗ 𝑓𝑒𝑥𝑡 𝑖 (1)
each step and usefulness of these architectures in different
fields and their applications. In the next chapter we go 𝑐𝑜𝑛𝑣𝑜𝑢𝑡 = ((𝑣𝑖 − 2𝑝𝑎𝑑𝑑 + 𝑝𝑖 )⁄𝑠𝑡𝑟 ) + 1 (2)
through the methodology opted by us and the design details
i.e., our methodology and the implementation of our method 1
𝑒𝑟𝑟𝑐𝑎𝑙𝑐 = ( ) ∑𝑁
𝑖=0 ∇𝑖 × 𝑙𝑜𝑠𝑠𝑓𝑢𝑛𝑐 (𝑖𝑚𝑔𝑖 , 𝐶𝑛𝑛𝑜𝑢𝑡 𝑖 , 𝑤𝑔𝑡𝑖 )
into a working prototype. In the next chapter we discuss 𝑁
about the reliability and usefulness of our model, the results (3)
both in simulation and in real-time and the comparison of
our design with the other existing models.

II. LITERATURE SURVEY

The backend or the main architecture behind the YOLO
algorithm is CNN (convolutional neural network). CNNs are
specialized in the detection and classification of image data.
Similar to other neural networks, CNNs have three types of
layers- input layer, hidden layer(s) and an output layer.
Fig. 1: Basic CNN architecture
The layers of the network consist of a number of
neurons each connected to neurons in the successive layer.
These connections between neurons are known as synapse YOLO is structured in the same way but the output
and each synapse is associated with a particular layer of the YOLO consists of a fully connected
weight(wgt). The CNN extracts the main features of the convolutional layer. Where a CNN implements the sliding
given image with the help of feature extractors(fext). With window technique for the detection of objects, the YOLO
the help of zero padding(padd), the input image is padded algorithm directly predicts the class probabilities and
with additional zeros to overcome the shrinkage of the bounding boxes of the object based on the position of
image which occurs during convolution. These feature midpoint (center) of the object in a given grid of the input
extractors(kernels) make up a feature map with all the image. [25]
features of the given input image. All these operations are
performed in the form of matrices. A ReLu activation
function is used for convolution. A pooling layer,
conventionally a max pooling layer is used to extract the
distinguishing features of the image. This is for a single
operation with convolution layer, and max-pooling layer.
The final layers of a conventional CNN involve a flattening
layer (to reduce the higher dimension data to a single
dimension) and a fully connected layer. Normally a Fig.2: Tiny_Yolov3 architecture

National Conference On Advances in Electronics Signal Processing and Communications (AESPC-2020)

Volume 6 Issue 3, JUNE 2020
 INTERNATIONAL JOURNAL OF RESEARCH IN TECHNOLOGY AND MANAGEMENT (IJRTM)
ISSN 2454-6240
www.ijrtm.com
Tiny_yolov3 is proved to be the fastest of all the yolo in a single network. Yolo considers the detection process as
versions developed so far. The working of tiny_yolov3 is a regression problemIt is very fast as the classification and
similar to that of a conventional yolo model. The YOLO prediction of the object is all done by a single evaluating
algorithm divides the image into a grid of dimensions T×T. layer. RCNN on the other hand classifies the objects in two
The prediction or output (yolo_out) of the yolo algorithm is stages i.e., the classification of the object (category)and the
a function of the probability(pobj) or confidence that an location of the object and bounding boxes (regression).
object exists in a given grid square, the position of midpoint Since it’s a two-stage detection, it is very much slower than
(center) of the object (midx, midy) in the grid square, the YOLO. An improved version of R-CNN known as Fast R-
height (boxh) and width (boxw) of the bounding box and the CNN is also developed which is faster than a regular R-
class probabilities (class1, class2, class3……classN). The CNN but still slower than YOLO[27]
prediction of the object depends upon the probability of an In UAV (Unmanned aerial Vehicle) network, there
object existing(pobj) in a given square of a grid. If there is no is a large background noise from the recorded or collected
object, then pobj becomes zero and all other quantities video images and also the dimensions of the vehicle are
become don’t cares (?). scaled so much that it is harder to detect a vehicle from the
video footage.This problem can be solved by the yolo
𝑦𝑜𝑙𝑜_𝑜𝑢𝑡 = 𝑓(𝑝𝑜𝑏𝑗 , 𝑚𝑖𝑑𝑥 , 𝑚𝑖𝑑𝑦 , 𝑏𝑜𝑥ℎ , 𝑏𝑜𝑥𝑤 , 𝑐𝑙𝑎𝑠𝑠1, algorithm as it not only detects the smaller objects through
𝑐𝑙𝑎𝑠𝑠2, 𝑐𝑙𝑎𝑠𝑠3, … . , 𝑐𝑙𝑎𝑠𝑠𝑁) (4) clustering the estimated sizes but also helps the system to
extract the smaller features of the vehicle through optimal
Each square of the grid evaluates only one set of pooling strategy and dense topology.[28]
class probabilities at a time so the number of objects
detected by a single grid is one. To detect more objects, we
III. PROPOSED METHOD
use the help of anchors. Anchors are useful when more than
a single object is present or assigned to a given square of the In our proposed method we use the varied version
grid. Based on the shape of the object, the model can of the yolo algorithm which is tiny_yolov3. This algorithm
classify and detect the objects in the given input. has only nine convolutional layers as opposed to the basic
A number of predictions are made for the same yolo algorithm which has twenty-one convolutional layers.
object from which only one is very close to the ground truth. We take in the disease’s physical characteristics as the input.
To select an accurate prediction, we use a method called non These are processed and given to our model which in turn is
max suppression (NMS). Let us assume N predictions have trained to detect and classify the disease. The output of our
been made, each with an individual confidence of Cnd and a proposed method is a disease detection and classification
threshold value of Th. Let the output of NMS method be system.
finalout. Firstly, a prediction with a high value of Cnd is
selected and assumed to be finalout. This prediction is then
compared with another prediction i.e., IOU (intersection
over union) of these two predictions is calculated.IOU is a
method used to calculate the overlapping area between two
predictions. This process is repeated for all the predictions. Fig. 3: Our proposed method
If the value is greater than the threshold value ‘Th’, then that
prediction is removed. Normally, a threshold of 0.5 is We prepared our dataset from the Kaggle dataset
selected. This is repeated until only one prediction remains repository, UCI machine learning repository and the
which is finalout. In this way, an object is detected only once dataset provided to us by the SmartBridge organization.
by the model. The dataset consists of 2 classes (False Smut and Leaf
Blast) of images with 500 images per each class. These
𝐼𝑂𝑈 = 𝐴𝑟𝑒𝑎 𝑜𝑓 𝑖𝑛𝑡𝑒𝑟𝑠𝑒𝑐𝑡𝑖𝑜𝑛⁄𝐴𝑟𝑒𝑎 𝑜𝑓 𝑢𝑛𝑖𝑜𝑛 (5) images are again augmented to make a total of 2000
images out of which 1600 images are used as training
set; 200 images are used as test set and the remaining
Yolo is extremely fast, since, unlike other are used as validation set.
classification models, it uses only one evaluation layer. We divided the training set into a batch-size of 64.
Therefore, it is used in various real time applications such as We set the learning rate to 0.001. A decay of 0.0005 is
traffic monitoring and pedestrian detection. A Yolo_v2 employed so the weights don’t take on large values.
network is used. Here the number of layers of Yolo_v2 is The random parameter is set to one so as to resize the
changes from 16 to 12 layers in the route layer of the input images to various sizes and generalize the objects
passthrough layer which results in an improvised network thus augmenting data. The parameters jitter, angle, hue
YOLO-R. This network is then trained to detect the and saturation are also employed for the data
pedestrians [26]. augmentation. The thresh parameter is the IOU value
Yolo is also used in real time face detection. The which is set to 0.5 and the ignore_thresh to 0.7. This
algorithms mainly used for this application are YOLO and ensures that a valid prediction is selected from the
R-CNN (regional CNN).The differences in the architecture various other predictions during non-max suppression.
between these two networks is that in the yolo model, the Six anchors are initialized to detect the different
training, extraction of details and classification are all done diseases occurring in different sizes. Since the diseases
may occur in many forms, it is always a good practice
National Conference On Advances in Electronics Signal Processing and Communications (AESPC-2020)
Volume 6 Issue 3, JUNE 2020
 INTERNATIONAL JOURNAL OF RESEARCH IN TECHNOLOGY AND MANAGEMENT (IJRTM)
ISSN 2454-6240
www.ijrtm.com
to initialize a decent number of anchor boxes such that
the detection (bounding box) is always accurate. Note
that initialization of a lot of bounding boxes may also
cause to predict false positives where in the model
confuses one object (disease) with another. So, the
number of anchor boxes should not be less as it may
neglect few objects and not too high as to predict false
positives.
We used the ‘leaky’ activation function for the
initial convolution layers. We also perform zero
padding of size (1,1) to avoid the loss in resolution. For
the last convolutional layer, we use the ‘linear’
activation function. In this way, the prediction of the
disease becomes a regression problem as is viewed by
our tiny_yolov3 model.
We used a remote Linux server for developing our Fig. 5 (a) Detection of False Smut disease in Rice crop using
YOLO algorithm-1
model along with DarkNet. DarkNet framework is
helpful in running faster computations with support of
GPU. Computations run a lot faster on GPUs than
conventional CPUs. It is written in CUDA and C and
supports all CUDA supported GPUs. In order to train
the model, we used Jetson nano (CUDA supported) as
our GPU platform, so as to reduce the training time. We
take the help of CUDA and OpenCV and compile the
darknet to start our training. It’s 128 CUDA cores
significantly increase the speed of computing.After the
training, the weights are frozen and stored for further
testing and classification.

IV. RESULTS AND DISCUSSIONS

The model is thoroughly designed to detect even the Fig. 5 (b)Detection of False Smut disease in Rice crop using
tiniest of the details. We up-sample the given input thus YOLO algorithm-2
profoundly extracting the features. This is done because the
image loses its resolution after successive convolutions and
thus some of the important features might be lost. Up-
sampling makes sure that the model covers the depth of the
image and each and every minute detail is extracted. We
achieved an accuracy of 98.92% which is by far the highest
for any state-of-the-art object detection algorithm. The
detection can be done onboard the Jetson nano board with
the help of raspberry-pi camera.
The detection can also be performed from a PC or
laptop with the help of python and OpenCV. The webcam
can be used to detect the disease in real time. With the help
of libraries such as Tkinter, we can detect the disease from
an image of the diseased plant. In this way, it’s very flexible
to detect the disease. Though the mode of detection is
changing, the accuracy of the model remains the same thus
making it a reliable object detection model.

Fig. 6 (a)Detection of Leaf Blast disease in rice crop using Yolo

algorithm-1

National Conference On Advances in Electronics Signal Processing and Communications (AESPC-2020)

Volume 6 Issue 3, JUNE 2020
 INTERNATIONAL JOURNAL OF RESEARCH IN TECHNOLOGY AND MANAGEMENT (IJRTM)
ISSN 2454-6240
www.ijrtm.com
REFERENCES
[1]. Rao, A. N., Wani,S. P., Ramesha ,M. S. and Ladha, J. K. 2017.
"Rice production systems," Rice Production Worldwide
(February 2017), 185-205.
[2]. Thorburn, C. 2015. "The rise and demise of integrated pest
management in rice in Indonesia," Insects(April 2015), 381–
408.
[3]. Garcia, J and Barbedo, A.2013. "Digital image processing
techniques for detecting, quantifying and classifying plant
diseases," Springer Plus, 2 (2013) 660.
[4]. Heinrichs, E. A. 2017 Retreived September 2017 from
ipmworld: https://ipmworld.umn.edu/heinrichs
[5]. Alam, M. Z., Crump,A. R., Haque, M. M., Islam, M. S.,
Hossain, E., Hasan S. B., Hasan ,S. B. and Hossain, M. S. 2016.
"Effects of integrated pest management on pest damage and
yield components in a rice agro-ecosystem in the Barisal Region
of Bangladesh," Environmental Science (March 2016)
[6]. Patil, J. K. and Kumar,R. 2011. "Advances in image processing
for detection of plant diseases," Journal of Advanced
Fig. 6 (b)Detection of Leaf Blast disease in rice crop
Bioinformatics Applications and Research, 2(2)(2011) 135-141.
using Yolo algorithm-2
[7]. Headey, D. and Hoddinott, J. 2016. "Agriculture, nutrition and
the green revolution in Bangladesh". Agricultural Systems 149
(Nov. 2016) 122-131.
[8]. Yamsake, S. and Resosudarmo, B. P. 2008. "Does sending
farmers back to school have an impact? revisiting the issue,"
Blackwell Publishing Asia(2008).
TABLE I. COMPARISON WITH OTHER EXISTING METHODS[29] [9]. Berg H. D., and Jiggins,J. 2007. "Investing in farmers—the
impacts of farmer field schools in relation to integrated pest
Model Accuracy management," World Development, 35(4)(2007) 663-686.
[10]. Hong-xing ,X., Ya-jun Y., Yan-hui ,L., Xu-song, Z., Junce ,T.
VGG16 97.12 and Feng-xiang, L.2017. "Sustainable management of rice insect
pests by non-chemical-insecticide technologies in China," Rice
MobileNetv2 96.12 Science, 24(2)(2017) 61-72.
[11]. Bajwa ,W. I. and Kogan, M. 2017. "Internet-based IPM
NasNet Mobile 96.95 informatics and decision Support," Integrated Plant Protection
Center (IPPC), Retrieved September 2017 from ipmworld:
Inception V3 96.37 https://ipmworld.umn.edu/bajwa.
[12]. MacRae, I. V. 2017. "IPM World," University of Minnesota,
Tiny_YoloV3* 98.92
Colorado State University, Retrieved September 2017 from
ipmworld : https://ipmworld.umn.edu/macrae-ipm.
[13]. Azfar,S., Nadeem ,A., and Basit ,A. 2015. "Pest detection and
We can see that our model has the highest accuracy control techniques using wireless sensor network", Journal of
compared to other existing models. Also, our model is faster Entomology and Zoology Studies, 3(2)(2015) 92- 99.
with a refresh rate of about 45 fps whereas the refresh rate [14]. Revathi, P. and Hemalatha, M. 2012. "Classification of cotton
for other models is restricted around 15-20fps. Though the leaf spot diseases using image processing edge detection
techniques," Emerging Trends in Science, Engineering and
localization errors are more in yolo model, it is faster than Technology
the other models which makes it the most reliable algorithm [15]. Donatelli, M., Magarey, R.D., Bregaglio, S., Willocquet L.,
suited for the problem. Whish, J.P.M., and Savary, S. 2017. "Modelling the impacts of
pests and diseases on agricultural systems," Agricultural
Systems, 155(2017) 213-224.
V. CONCLUSION [16]. Hira, J. K. and Din, S. 2015. "Design and development of
software for disease detection and management of leafy
In this paper, we proposed a deep learning model based on vegetables," International Journal of Innovative Research in
tiny_yolov3 algorithm to detect the rice crop diseases. The Computer and Communication Engineering, 3(6)(2015) 2320-
9798.
model is very fast and hence it can be applied in real time [17]. Wang ,K., Zhang, S., Wang ,Z., Liu, Z. and Yang,F. 2013.
for the detection and classification of the rice disease. This "Mobile smart device-based vegetable disease and insect pest
is helpful in detecting the diseases in the early stages so the recognition method," Yang Intelligent Automation & Soft
farmers can take preventive measures and protect the crop Computing, 19(3) (2013).
[18]. Phadikar, S. and Sil, J. 2009. "Rice disease identification using
from destruction, thus maximizing the annual yield. The pattern recognition techniques," Computer and Information
resolution of the camera doesn’t matter as the model is Technology, 2009.
designed to detect even the smallest of the features and thus [19]. Gohokar ,V. V., Mundada, R., "Detection and Classification of
obsoleting the need for high resolution cameras. Future Pests in Greenhouse Using Image Processing," Journal of
Electronics and Communication Engineering , 5(6)(2013), 57-
works include a model with more classes of diseases. 63.
Adding extra classes to the model results in a fully [20]. Singh , V. and Misrab, A.K. 2016. "Detection of plant leaf
operational rice crop disease classifier where it can detect diseases using image segmentation and soft computing
and classify all the diseases.The algorithm achieved an techniques", Information Processing in Agriculture, 4(2016),
41-49.
accuracy of 98.92%. Since we designed our model based on [21]. Huddar, S. R., Gowri, S., Keerthana, K., Vasanthi, S. and
tiny_yolov3, it can be further applied to detect pests as well Rupanagudi, S. R. 2012. "Novel algorithm for segmentation and
making it flexible for development of new models. automatic identification of pests on plants using image
processing," Computing Communication & Networking
Technologies , (2012).
National Conference On Advances in Electronics Signal Processing and Communications (AESPC-2020)
Volume 6 Issue 3, JUNE 2020
 INTERNATIONAL JOURNAL OF RESEARCH IN TECHNOLOGY AND MANAGEMENT (IJRTM)
ISSN 2454-6240
www.ijrtm.com
[22]. Kawasaki, Y., Uga, H., Kagiwada, S. and Iyatom, H. 2015. [28]. Z. Xu, H. Shi, N. Li, C. Xiang and H. Zhou, "Vehicle Detection
"Basic study of automated diagnosis of viral plant diseases Under UAV Based on Optimal Dense YOLO Method," 2018
using convolutional neural networks," Advances in Visual 5th International Conference on Systems and Informatics
Computing, (2015),638-645. (ICSAI), Nanjing, 2018, pp. 407-411, doi:
[23]. Sladojevic, S., Arsenovic, M., Anderla, A., Culibrk, D. and 10.1109/ICSAI.2018.8599403.
Stefanovic, D.2016. "Deep neural networks based recognition of [29]. Rahman, C. R., Arko, P. S., Ali, M. E., Iqbal Khan, M. A.,
plant diseases by leaf image classification," Computational Apon, S. H., Nowrin, F., & Wasif, A. (2020). Identification and
Intelligence and Neuroscience, (2016), 11. recognition of rice diseases and pests using convolutional neural
[24]. Yao, Q., Chen, G., Wang, Z., Zhang, C., Yang, B. and Tang, J. networks. Biosystems Engineering, 194, 112–120.
2017. "Automated detection and identification of whitebacked doi:10.1016/j.biosystemseng.2020.03.020
planthoppers in paddy fields using image processing," Journal
of Integrative Agriculture, 16(7)(2017), 1547-1557
[25]. You Only Look Once: Unified, Real-Time Object Detection
Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi
University of Washington, Allen Institute for AI, Facebook AI
Research
[26]. W. Lan, J. Dang, Y. Wang and S. Wang, "Pedestrian Detection
Based on YOLO Network Model," 2018 IEEE International
Conference on Mechatronics and Automation (ICMA),
Changchun, 2018, pp. 1547-1551, doi:
10.1109/ICMA.2018.8484698.
[27]. W. Yang and Z. Jiachun, "Real-time face detection based on
YOLO," 2018 1st IEEE International Conference on Knowledge
Innovation and Invention (ICKII), Jeju, 2018, pp. 221-224, doi:
10.1109/ICKII.2018.8569109.

National Conference On Advances in Electronics Signal Processing and Communications (AESPC-2020)

Volume 6 Issue 3, JUNE 2020